Instructions to authors Aims and scope Physics Reports keeps the active physicist up-to-date on developments in a wide range of topics by publishing timely reviews which are more extensive than just literature surveys but normally less than a full monograph. Each Report deals with one speci"c subject. These reviews are specialist in nature but contain enough introductory material to make the main points intelligible to a non-specialist. The reader will not only be able to distinguish important developments and trends but will also "nd a su$cient number of references to the original literature. Submission In principle, papers are written and submitted on the invitation of one of the Editors, although the Editors would be glad to receive suggestions. Proposals for review articles (approximately 500}1000 words) should be sent by the authors to one of the Editors listed below. The Editor will evaluate proposals on the basis of timeliness and relevance and inform the authors as soon as possible. All submitted papers are subject to a refereeing process. Editors J.V. ALLABY (Experimental high-energy physics), PPE Division, CERN, CH-1211 Geneva 23, Switzerland. E-mail:
[email protected] D.D. AWSCHALOM (Experimental condensed matter physics), Department of Physics, University of California, Santa Barbara, CA 93106, USA. E-mail:
[email protected] J.A. BAGGER (High-energy physics), Department of Physics, The Johns Hopkins University, 3400 North Charles Street, Baltimore MD 21218, USA. E-mail:
[email protected] C.W.J. BEENAKKER (Mesoscopic physics), Instituut}Lorentz, Universiteit Leiden, P.O. Box 9506, 2300 RA Leiden, The Netherlands. E-mail:
[email protected] E. BREZIN (Statistical physics and ,eld theory), Laboratoire de Physique Theorique, Ecole Normale Superieure, 24 rue Lhomond, 75231 Paris Cedex, France. E-mail:
[email protected] G.E. BROWN (Nuclear physics), Institute for Theoretical Physics, State University of New York at Stony Brook, Stony Brook, NY 11974, USA. E-mail:
[email protected] D.K. CAMPBELL (Non-linear dynamics), Dean, College of Engineering, Boston University, 44 Cummington Street, Boston, MA 02215, USA. E-mail:
[email protected] G. COMSA (Surfaces and thin ,lms), Institut fuK r Physikalische und Theoretische Chemie, UniversitaK t Bonn, Wegelerstrasse 12, D-53115 Bonn, Germany. E-mail:
[email protected] J. EICHLER (Atomic and molecular physics), Hahn-Meitner-Institut Berlin, Glienicker Strasse 100, 14109 Berlin, Germany. E-mail:
[email protected] T.F. GALLAGHER (Atomic and molecular physics), Department of Physics, University of Virginia, Charlottesville, VA 22901, USA. E-mail:
[email protected]
vi
Instructions to authors
M.P. KAMIONKOWSKI (Astrophysics), Theoretical Astrophysics 130-33, California Institute of Technology, 1200 East California Blvd., Pasadena, CA 91125, USA. E-mail:
[email protected] M.L. KLEIN (Soft condensed matter physics), Department of Chemistry, University of Pennsylvania, Philadelphia, PA 19104-6323, USA. E-mail:
[email protected] A.A. MARADUDIN (Condensed matter physics), Department of Physics, University of California, Irvine, CA 92717, USA. E-mail:
[email protected] D.L. MILLS (Condensed matter physics), Department of Physics, University of California, Irvine, CA 92717, USA. E-mail:
[email protected] R. PETRONZIO (High-energy physics), Dipartimento di Fisica, II Universita` di Roma } Tor Vergata, Via Orazio Riamondo, 00173 Rome, Italy. E-mail:
[email protected] S. PEYERIMHOFF (Molecular physics), Institute of Physical and Theoretical Chemistry, Wegelerstrasse 12, D-53115 Bonn, Germany. E-mail:
[email protected] I. PROCACCIA (Statistical mechanics), Department of Chemical Physics, Weizmann Institute of Science, Rehovot 76100, Israel. E-mail:
[email protected] E. SACKMANN (Biological physics), Physik-Department E22 (Biophysics Lab.), Technische UniversitaK t MuK nchen, D-85747 Garching, Germany. E-mail:
[email protected] A. SCHWIMMER (High-energy physics), Physics Department, Weizmann Institute of Science, Rehovot 76100, Israel. E-mail:
[email protected] R.N. SUDAN (Plasma physics), Laboratory of Plasma Studies, Cornell University, 369 Upson Hall, Ithaca, NY 14853-7501, USA. E-mail:
[email protected] W. WEISE (Physics of hadrons and nuclei), Institut fuK r Theoretische Physik, Physik Department, Technische UniversitaK t MuK nchen, James Franck Stra{e, D-85748 Garching, Germany. E-mail:
[email protected] Manuscript style guidelines Papers should be written in correct English. Authors with insu$cient command of the English language should seek linguistic advice. Manuscripts should be typed on one side of the paper, with double line spacing and a wide margin. The character size should be su$ciently large that all subscripts and superscripts in mathematical expressions are clearly legible. Please note that manuscripts should be accompanied by separate sheets containing: the title, authors' names and addresses, abstract, PACS codes and keywords, a table of contents, and a list of "gure captions and tables. } Address: The name, complete postal address, e-mail address, telephone and fax number of the corresponding author should be indicated on the manuscript. } Abstract: A short informative abstract not exceeding approximately 150 words is required. } PACS codes/keywords: Please supply one or more PACS-1999 classi"cation codes and up to 4 keywords of your own choice for indexing purposes. PACS is available online from our homepage (http://www.elsevier.com/locate/physrep). References. The list of references may be organized according to the number system or the name-year (Harvard) system. Number system: [1] M.J. Ablowitz, D.J. Kaup, A.C. Newell and H. Segur, The inverse scattering transform } Fourier analysis for nonlinear problems, Studies in Applied Mathematics 53 (1974) 249}315.
Instructions to authors
vii
[2] M. Abramowitz and I. Stegun, Handbook of Mathematical Functions (Dover, New York, 1965). [3] B. Ziegler, in: New Vistas in Electro-nuclear Physics, eds E.L. Tomusiak, H.S. Kaplan and E.T. Dressler (Plenum, New York, 1986) p. 293. A reference should not contain more than one article. Harvard system: Ablowitz, M.J., D.J. Kaup, A.C. Newell and H. Segur, 1974. The inverse scattering transform } Fourier analysis for nonlinear problems, Studies in Applied Mathematics 53, 249}315. Abramowitz, M. and I. Stegun, 1965, Handbook of Mathematical Functions (Dover, New York). Ziegler, B., 1986, in: New Vistas in Electro-nuclear Physics, eds E.L. Tomusiak, H.S. Kaplan and E.T. Dressler (Plenum, New York) p. 293. Ranking of references. The references in Physics Reports are ranked: crucial references are indicated by three asterisks, very important ones with two, and important references with one. Please indicate in your "nal version the ranking of the references with the asterisk system. Please use the asterisks sparingly: certainly not more than 15% of all references should be placed in either of the three categories. Formulas. Formulas should be typed or unambiguously written. Special care should be taken of those symbols which might cause confusion. Unusual symbols should be identi"ed in the margin the "rst time they occur. Equations should be numbered consecutively throughout the paper or per section, e.g., Eq. (15) or Eq. (2.5). Equations which are referred to should have a number; it is not necessary to number all equations. Figures and tables may be numbered the same way. Footnotes. Footnotes may be typed at the foot of the page where they are alluded to, or collected at the end of the paper on a separate sheet. Please do not mix footnotes with references. Figures. Each "gure should be submitted on a separate sheet labeled with the "gure number. Line diagrams should be original drawings or laser prints. Photographs should be contrasted originals, or high-resolution laserprints on glossy paper. Photocopies usually do not give good results. The size of the lettering should be proportionate to the details of the "gure so as to be legible after reduction. Original "gures will be returned to the author only if this is explicitly requested. Colour illustrations. Colour illustrations will be accepted if the use of colour is judged by the Editor to be essential for the presentation. Upon acceptance, the author will be asked to bear part of the extra cost involved in colour reproduction and printing. For details, contact the Publisher at the address below. After acceptance } Proofs: Proofs will be sent to the author by e-mail, 6}8 weeks after receipt of the manuscript. Please note that the proofs have been proofread by the Publisher and only a cursory check by the author is needed; we are unable to accept changes in, or additions to, the edited manuscript at this stage. Your proof corrections should be returned as soon as possible, preferably within two days of receipt by fax, courier or airmail. The Publisher may proceed with publication of no response is received. } Copyright transfer: The author(s) will receive a form with which they can transfer copyright of the article to the Publisher. This transfer will ensure the widest possible dissemination of information. LaTeX manuscripts The Publisher welcomes the receipt of an electronic version of your accepted manuscript (encoded in LATEX). If you have not already supplied the "nal, revised version of your article (on diskette) to the Journal Editor, you are requested herewith to send a "le with the text of the manuscript
viii
Instructions to authors
(after acceptance) directly to the Publisher by e-mail or on diskette to the address given below. Please note that no deviations from the version accepted by the Editor of the journal are permissible without the prior and explicit approval by the Editor. Such changes should be clearly indicated on an accompanying printout of the "le. Files sent via electronic mail should be accompanied by a clear identi"cation of the article (name of journal, editor's reference number) in the &&subject "eld'' of the e-mail message. LATEX articles should use the Elsevier document class &&elsart'', or alternatively the standard document class &&article''. The Elsevier LATEX package (including detailed instructions for LATEX preparation) can be obtained from http://www.elsevier.com/locate/latex. The elsart package consists of the "les: ascii.tab (ASCII table), elsart.cls (use this "le if you are using LATEX2e, the current version of LATEX), elsart.sty and elsart12.sty (use these two "les if you are using LATEX2.09, the previous version of LATEX), instraut.dvi and/or instraut.ps (instruction booklet), readme. Author bene5ts } Free o+prints. For regular articles, the joint authors will receive 25 o!prints free of charge of the journal issue containing their contribution; additional copies may be ordered at a reduced rate. } Discount. Contributors to Elsevier Science journals are entitled to a 30% discount on all Elsevier Science books. } Contents Alert. Physics Reports is included in Elsevier's pre-publication service Contents Alert. Author enquiries For enquiries relating to the submission of articles (including electronic submission), the status of accepted articles through our Online Article Status Information System (OASIS), author Frequently Asked Questions and any other enquiries relating to Elsevier Science, please consult http://www.elsevier.com/locate/authors/ For speci"c enquiries on the preparation of electronic artwork, consult http://www.elsevier.com/ locate/authorartwork/ Contact details for questions arising after acceptance of an article, especially those relating to proofs, are provided when an article is accepted for publication.
THE PDF APPROACH TO TURBULENT POLYDISPERSED TWO-PHASE FLOWS
Jean-Pierre MINIER, Eric PEIRANO ElectriciteH de France, Div. R&D, MFTT, 6 Quai Watier, 78400 Chatou, France Energy Conversion Department, Chalmes University of Technology, S-41296 GoK teborg, Sweden
AMSTERDAM } LONDON } NEW YORK } OXFORD } PARIS } SHANNON } TOKYO
Physics Reports 352 (2001) 1–214
The pdf approach to turbulent polydispersed two-phase ows Jean-Pierre Miniera ; ∗ , Eric Peiranob b
a Electricite de France, Div. R&D, MFTT, 6 Quai Watier, 78400 Chatou, France Energy Conversion Department, Chalmers University of Technology, S-41296 G-oteborg, Sweden
Received December 2000; editor : I: Procaccia
Contents 1. Introduction 1.1. Two-phase ow regimes 1.2. An industrial example of dispersed two-phase ows 1.3. Mathematical and physical approach 1.4. Description of the contents 2. Mathematical background on stochastic processes 2.1. Random variables 2.2. Stochastic processes 2.3. Markov processes 2.4. Key Markov processes 2.5. General Chapman–Kolmogorov equations 2.6. Stochastic di8erential equations and di8usion processes 2.7. Stochastic calculus 2.8. Langevin and Fokker–Planck equations 2.9. The probabilistic interpretation of PDEs 2.10. A word on numerical schemes 3. Hierarchy of pdf descriptions
3 4 5 7 11 12 13 15 16 17 19 22 24 25 26 27 28
3.1. Complete and reduced pdf equations 3.2. BBGKY hierarchy 3.3. Hierarchy between state vectors 4. Stochastic di8usion processes for modelling purposes 4.1. The shift from an ODE to a SDE 4.2. Modelling principles 4.3. Example for typical stochastic models 5. The physics of turbulence 5.1. The turbulence problem 5.2. Characteristic scales 5.3. Kolmogorov theory 5.4. Di@culties and reAnements 5.5. Experimental and numerical results 5.6. SimpliAed images of turbulence and Lagrangian models 5.7. Closing remarks 6. One-point pdf models in single-phase turbulence 6.1. Motivation and basic ideas 6.2. Coarse-grained description and stochastic modelling
∗
Corresponding author. Tel.: +33-1-30-87-71-40; fax: +33-1-30-87-79-16. E-mail addresses:
[email protected] (J.-P. Minier),
[email protected] (E. Peirano). c 2001 Elsevier Science B.V. All rights reserved. 0370-1573/01/$ - see front matter PII: S 0 3 7 0 - 1 5 7 3 ( 0 1 ) 0 0 0 1 1 - 4
29 30 32 34 35 36 38 40 41 42 46 51 55 59 64 65 66 66
2
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
6.3. Relations to classical approaches 6.4. Probabilistic description of continuous Aelds 6.5. Choice of the pdf description 6.6. Present one-point models 6.7. Mean Aeld equations 6.8. Physical and information contents 6.9. Numerical examples 7. One-point particle pdf models in two-phase ows 7.1. Fundamental equations and modelling approaches 7.2. Interest of the pdf approach 7.3. Choice of the pdf description 7.4. Present models 7.5. Properties of present class of models 7.6. Numerical examples and typical simulations
67 69 75 77 80 84 88 99 100 105 106 108 120 135
8. Two-point uid–particle pdf models in dispersed two-phase ows 8.1. Motivations and basic ideas 8.2. Probabilistic description of dispersed two-phase ows 8.3. Choice of the pdf description 8.4. Present ‘two-point’ models 8.5. Mean Aeld equations 8.6. Concluding remarks 9. Summary and propositions for new developments 9.1. Di@culties with conventional approaches and interest of a pdf description 9.2. Assessment of current modelling state 9.3. Open issues and suggestions References
148 149 149 161 164 169 181 181 181 184 186 209
Abstract The purpose of this paper is to develop a probabilistic approach to turbulent polydispersed two-phase ows. The two-phase ows considered are composed of a continuous phase, which is a turbulent uid, and a dispersed phase, which represents an ensemble of discrete particles (solid particles, droplets or bubbles). Gathering the di@culties of turbulent ows and of particle motion, the challenge is to work out a general modelling approach that meets three requirements: to treat accurately the physically relevant phenomena, to provide enough information to address issues of complex physics (combustion, polydispersed particle ows, : : :) and to remain tractable for general non-homogeneous ows. The present probabilistic approach models the statistical dynamics of the system and consists in simulating the joint probability density function (pdf) of a number of uid and discrete particle properties. A new point is that both the uid and the particles are included in the pdf description. The derivation of the joint pdf model for the uid and for the discrete particles is worked out in several steps. The mathematical properties of stochastic processes are Arst recalled. The various hierarchies of pdf descriptions are detailed and the physical principles that are used in the construction of the models are explained. The Lagrangian one-particle probabilistic description is developed Arst for the uid alone, then for the discrete particles and Anally for the joint uid and particle turbulent systems. In the case of the probabilistic description for the uid alone or for the discrete particles alone, numerical computations are presented and discussed to illustrate how the method works in practice and the kind of information that can be extracted from it. Comments on the current modelling state and propositions for future investigations which try to link the present c 2001 Elsevier Science B.V. All work with other ideas in physics are made at the end of the paper. rights reserved. PACS: 47.27.Eq; 47.55.Kf; 02.40.+j; 02.50.Ey Keywords: Turbulence; Two-phase ows; Probability density function; Stochastic process
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
3
1. Introduction In March 1999, Reviews of Modern Physics issued a special volume, for the commemoration of its 100th birthday, which discussed historical developments and gave a general outlook on a wide range of physical questions. Numerous articles, written by world experts and often major contributors to their Aelds, provided an overview of past achievements and of the current state of each domain. Apart from the Aelds which have traditionally formed the main core of theoretical physics (quantum theory, particle physics, relativity, astrophysics, etc.) the selection of other subjects (such as uid turbulence, granular matter, soft matter, biological physics) which also found their place in this prestigious assembly is an indication of the interest for the issues raised by these subjects. This is also an indication that improved physical understanding is needed to bring these subjects to a more mature state. In the section devoted to statistical physics, two reviews, written by Sreenivasan and De Gennes, respectively, discussed separately the present understanding of uid turbulence [1] and of granular matter [2] (say, the behaviour of non-Brownian small solid particles). Broadly speaking, both are subjects where the basic equations (for example, the Navier–Stokes equations) or the elementary behaviour (for instance, how two grains interact) may be believed to be known, but where the issue is to understand the complicated and collective behaviour of a large number of interacting degrees of freedom. Both represent problems at a human-size level. They are actually everyday-life concerns and could, at Arst, have been thought to be mere engineering problems. They are indeed engineering problems, but even if only approximate results or rough estimates are sought, this often requires a clear and precise physical understanding of the important phenomena at play. There is another interesting domain which is simply obtained when the two di@culties are mixed: the case of turbulent dispersed two-phase ows. An easy way to picture this is to imagine dealing with granular matter but embedded in a turbulent ow. These ows are of crucial importance in a large variety of industrial problems. Yet, they have not received the same attention as turbulence or as granular matter. As a consequence, physical understanding remains limited and appears to be scarce compared to each of the separate sub-cases, uid turbulence in the absence of particles, and granular matter in the absence of any underlying or interstitial uid. The purpose of the present work is to discuss some of the physical issues involved in two-phase ows and to put forward a probabilistic formalism that can bridge the gap between physical understanding of basic phenomena and practical simulations. That middle-road approach is that of a modeller, where one invents a model, which has simpliAed rules compared to the real phenomena, and which is used to simulate the overall and collective behaviour of a complex system. The question is therefore whether the model contains the right ‘physics’ (thus the need to understand clearly the important phenomena) and then how to reach an acceptable compromise between the simplicity of the model versus its physical realism (thus the need of an appropriate formalism). Before going into the details of the approach followed in this work, a clear deAnition of two-phase ows and particularly of dispersed two-phase ows must be given. Secondly, a better idea of their importance in natural and industrial situations as well as an outline of the modelling issues involved must be provided. Introducing these notions is perhaps best achieved through typical examples. Dispersed two-phase ows occur in many natural phenomena. They are met for example in fogs, in water sprays, in smokes, when desert sand is carried away or
4
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 1. Di8erent two-phase ow regimes in a heat exchanger.
sediments, or (to provide a more vivid image) when an erupting volcano billows around smoke and various particles. They are increasingly important in environmental problems when one or several species (not necessarily pollutants) are dispersed in a turbulent atmosphere. Nevertheless, in the following two sections and to complement these Arst examples, we introduce dispersed two-phase ows and discuss the modelling questions through industrial examples. 1.1. Two-phase ;ow regimes As it transpires from their name, two-phase ows are encountered when two non-miscible phases coexist. Depending on the form of the interface between the two media, di8erent regimes can be found. This is illustrated in Fig. 1 which shows a range of regimes for the case of a boiling liquid (for example water) in a classical heat exchanger. At the bottom of the tube, the liquid has not yet started to boil and we have a single-phase turbulent ow. When nucleation starts at the walls, bubbles can be found as separate inclusions within the liquid (bubbly ows). Then, as more vapour is created we go through the so-called slug and plug regimes where vapour occupies a more important volumetric fraction. Then, as the liquid continues to boil, we And the annular regime with a thin liquid layer at the walls and a central vapour ow with small droplets carried by the vapour. Other regimes can also be found when horizontal channels are considered, but their detailed description is outside the scope of the present article. The wide variety of regimes, merely outlined above, is typical of immiscible liquid–gas or liquid–liquid ows since the interface can be deformed. Two of these regimes (the bubbly and annular regimes) are characterized by the presence of one phase, either liquid or vapour, as separate inclusions embedded in the other phase. These are two examples of what is deAned as dispersed turbulent two-phase ows, where one phase (called the continuous phase) is a continuum and the other phase (called the dispersed phase) appears as separate inclusions dispersed within the continuous one, assumed here to be a turbulent uid. When the dispersed phase is characterized by a distribution in size, one speaks of a polydispersed turbulent two-phase ;ows. The dispersed regime (either mono or polydispersed) is of Arst importance in most cases. It is always found when the dispersed phase is made up by solid particles (solid particles in a gas or a liquid turbulent ow). It is often found for a liquid dispersed as separate droplets in a gas ow (sprays for example) or for two immiscible liquids where one liquid is dispersed in the other liquid. Indeed, the dispersion of one phase within another one increases considerably the surface of the separating interface
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
5
Fig. 2. A circulating uidized bed combustor.
and thus allows better mass and energy transfer between the two phases. These higher transfer rates explain that the dispersed regime is preferable. In the present work, we limit ourselves exclusively to the dispersed regime and we will talk of a uid (the continuous phase) and of discrete particles (which can represent either solid particles, bubbles or droplets). In most of the problems encountered, the dispersed particles have a distribution in size. Since the polydispersed case obviously contains the monodispersed situation as a simple sub-case, we will consider the realistic problem of polydispersed two-phase ows. 1.2. An industrial example of dispersed two-phase ;ows The limitation to the regime of dispersed two-phase ows is of course a simpliAcation with respect to general two-phase ows. Yet, the range of problems remains large, and each of these problems is di@cult. What are the main problems and what are the key issues? To provide some answers to that question, it is perhaps better to describe a relevant industrial example, circulating uidized bed (CFB) boilers. This is an industrial process for thermal energy generation. A sketch of a typical unit is displayed in Fig. 2. In a conAned domain (the combustion chamber), solids (inert sand and solid fuel or coal particles with a size distribution ranging from 100 m to 1 mm and an average density of order of magnitude 1000 kg m−3 ) are transported vertically by a gas (injected at the bottom) through the combustion chamber. The solids are captured at the exit by a separator (usually a cyclone), and reintroduced near the bottom of the combustion chamber, whereas the gas leaves the cyclone through an outlet duct. The solid particles are therefore
6
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
recycled to the combustion chamber and some particles may thus perform several loops in the process. This solid circulation is the key factor of the whole process (thus its name). It ensures an approximate homogeneous temperature within the combustion chamber which can thus be chosen as an optimum between the e@ciency of the process and the formation of noxious pollutants (the resulting low level of emission is one of the strong points of CFB units). It represents also a central aspect that must be mastered if one is to expect a satisfactory performance of the overall process. The ow of the gas–solid mixture is non-stationary and non-homogeneous with a vertical distribution of the particle concentration inside the combustion chamber. This vertical distribution is characterized by three interacting zones (as a function of increasing height): (1) a bottom bed which has the characteristics of a bubbling bed (gas ows through the bed in the form of large structures) and the concentration of discrete particles is so high that particle–particle interaction is a dominant mechanism (particles collide and possibly slide against each other), (2) a splash zone with high clustering and back-mixing activity and (3) a transport zone which exhibits a core/wall-layer structure (particles are entrained upwards in the core and fall down along the walls in the form of a thin boundary layer). In these regions large scale spatial inhomogeneities in the discrete particle concentration Aeld can be observed and for the gas large scale instabilities (pseudo-like turbulence) are present. In regions (2) and (3), particle loading (the local instantaneous ratio between the weight of particles and the weight of gas) is high enough so that turbulence is modulated and possibly modiAed by the presence of the particles. In addition particle segregation can be observed, that is to say the mean particle diameter decreases with height and large particles tend to migrate to the boundary layers. At the exit of the chamber, the particle-laden ow enters the cyclone(s). Cyclones are used here as separators (separating the solid coal particles from the gas in order to recycle them) and are key elements of the whole process. Indeed, should they fail to ensure a proper separation and consequently a proper particle recirculation, the whole process would not be able to run correctly. Cyclone performances are quantiAed by their e@ciency curve which is the fraction of solid particles being collected (and thus recycled) as a function of the particle diameters. It is important to be aware that cyclone e@ciencies are due to the complicated swirling motions and gas ow patterns within the cyclone and not to external forces such as gravity. In other words, both within the combustion chamber and within the cyclone separators, satisfactory performances of a CFB process are ensured by the local hydrodynamics of the two-phase ows rather than by external monitoring. In particular, a key parameter for a good functioning of a CFB boiler is the particle size distribution, to ensure, for example, suitable particle spatial distribution and residence time in the combustion chamber. It is mainly controlled, for small diameters, by the collection e@ciency of the cyclone, and for large diameters, by the characteristics of the fuel particles. This is, after all, an engineering problem. Numerous industrial or engineering problems may also involve di@cult questions. In some situations, complex questions or issues may become less relevant or secondary if engineers apply a high-enough ‘margin coe@cient’. This easy way around theoretical issues may lead researchers to believe that only clever or astute Axing or tinkering is needed. However, from the above outline of the CFB process, it is clear that this is not the case here, since the overall performance is a result of the local hydrodynamics throughout the process.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
7
In summary, the physics of turbulent gas-solid ows (in the case of circulating uidised bed boilers) contains a large spectrum of problems. Three main themes emerge, namely: (A) polydispersed two-phase ows (there is a range of particle diameters rather than a single value), (B) combustion, either within reactive gas ows or of the solid particles (as in the case of fuel particles within the CFB boiler), (C) turbulence, which is the central and common issue. Of course, these three main categories overlap and a wide range of sub-categories or classes of problems can be enumerated. For example, this concerns: • reactive ows (heterogeneous and homogeneous combustion), • particle dispersion (and di8usion of combustion gases), • turbulence modulation, possibly modiAcation of its nature, by the presence and motion of
solid particles embedded in the turbulent ow,
• particle–particle interactions (short- and long duration collisions), • swirling two-phase ows (selectivity curve of the cyclone) and • particle segregation, etc.
Other industrial and practical needs involving two-phase ows, such as pollutant dispersion in the atmosphere, combustion of fuel droplets within car engines, etc. would reveal the same picture and the same categories with an emphasis on one of these categories depending on the application. From the previous analysis, it appears that one has to built the link between uid-mechanics, classical mechanics, tribology, combustion and chemistry. The question to be answered is: how can we achieve this goal with a tractable formalism which has to be, in addition, suitable for numerical applications? 1.3. Mathematical and physical approach 1.3.1. The present objectives For the two-phase ows we consider in the present work, the central subject is turbulence. Turbulence of continuous-phase ows is further compounded by the e8ects of the discrete particles. Direct numerical simulations are possible in theory but are quite impossible in practice, at least for the typical examples described above. Most turbulent dispersed two-phase ows involve far too many degrees of freedom to be directly simulated. The issue is therefore to reduce the number of degrees of freedom to a tractable number and to come up with a contracted description. We are thus faced with a problem of non-equilibrium statistical physics where one tries to obtain a statistical model for a reduced number of degrees of freedom. Given the inherent complexity of the problems we have to deal with, the Arst choice is to limit ourselves to mean or average quantities. This is classical in most problems of statistical mechanics. In other words, we treat the solutions of the fundamental equations as random variables and we are interested in some statistics. Compared to the high complexity and to the beauty of the initial problem, this may look as limited and perhaps unchallenging objectives. However, it must be remembered that we are not dealing with only one problem, either single-phase turbulence, combustion or
8
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
two-phase ows, but typically with the three of them. Consequently, the aim is to develop a mathematical and physical approach that meets the following requirements: 1. the important physical phenomena, such as convection or mean pressure-gradient, are treated without approximation, 2. enough information is available to handle correctly issues of complex physics (combustion, polydispersed particles), 3. the resulting numerical model is tractable for non-homogeneous ows, 4. the model can be coupled to other approaches, either more fundamental or applied descriptions. It is clear from the third constraint that a compromise between detailed descriptions and practical applicability must be reached. The approach will necessarily be less fundamental compared to a number of theoretical approaches in turbulence for example [3–5]. It does not mean that the present choice contradicts more theoretically oriented works, but rather that the underlying objectives are somewhat di8erent. Actually, a satisfactory model respecting the three Arst requirements should easily beneAt from fundamental progress made in one of the three main themes (A), (B) or (C) listed above, which are of concern here. This is one of the reasons for the fourth item which also suggests that the approach can be used in relation to coarser descriptions in a multi-scale or multi-level simulation. The main challenge comes from the second constraint. It implies that we are not looking for a model or an approach which performs very well for only one theme but for an approach that can handle complex physics. For example, we are not looking for an approach which is perhaps the best candidate at the moment to simulate, say, isothermal incompressible single-phase turbulent ows but which requires new formalisms or new models for combustion. We are looking for an approach that can do a Ane job for single-phase turbulence and still be easily extended to handle combustion and dispersed two-phase ow issues within the same framework. This ‘engineering’ constraint has far-reaching consequences in terms of modelling choices and justiAes advanced methods. 1.3.2. Choice of the modelling approach Since we are mainly interested in some local mean statistics on a number of uid and discrete particle properties and since we have emphasized the practical side of the problem, it would seem that the path of least dissipation (for the modeller) consists in trying to derive directly a set of closed partial di8erential equations (PDE) for those mean variables. We refer to this approach as the moment approach or the conventional approach. It is indeed in line with the classical or conventional approaches in continuum mechanics where one handles Aelds which are solutions of some PDEs. For example, if we are interested in the mean uid velocity, we start from the Navier–Stokes equations (we consider here an incompressible ow for the sake of simplicity) 9Ui 9Ui 1 9P 92 Ui =− + 2 : (1) + Uj 9t 9xj 9xi 9 xj This equation contains all the information for the uid velocity. Then, following the classical approach we apply to this equation an averaging operator (the nature of this averaging operator, be it the Reynolds average or a spatial Alter, does not change the present point so we use
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
9
here the Reynolds decomposition while comments on spatial Altering will be developed in Section 9), written :. We obtain the unclosed PDE directly for the variable of interest U 9ui uj 1 9P 9Ui 9Ui =− + QUi − : + Uj 9t 9xj 9xi 9xj
(2)
This open equation has to be closed by resorting to a constitutive relation giving the unknown quantity, here ui uj (the Reynolds stress tensor), as a function of known variables, here the mean velocity. If we use vocabulary from statistical physics, we can say that the conventional approach is a macroscopic approach where one tries to obtain the macroscopic laws through closure relations which are written directly at the macroscopic level. If an acceptable macroscopic constitutive relation can be found, then this route is certainly the most cost-e8ective one since we explicitly calculate only what we want and nothing more. However, the success of this macroscopic approach hinges on the possibility to express unclosed terms through macroscopic laws. If such macroscopic relations cannot be explicitly written or involve far too drastic assumptions to yield acceptable results, then the conventional approach fails. Two typical examples of such problems are provided by the reactive source terms which enter the equations of single-phase turbulent combustion with Anite-rate chemistry and by the existence of a range of particle diameters (a polydispersed two-phase ow). Both issues will be explained in detail in Section 6 for the former and in Section 7 for the latter. In each case, one has to express the average value of a complicated function of some instantaneous variables (the uid instantaneous species mass fractions or the particle instantaneous diameters), whereas the conventional approach can only provide information on the Arst moments (usually the Arst two moments). We are faced with the problem of having to express a quantity such as S( ), where S is a complicated function of some scalar , in terms of the available information, usually limited to or 2 . This results in an intractable problem and more information is needed to address these typical issues of complex physics. In other words, even if we are interested mainly in estimates of macroscopic quantities, an advanced method providing more detailed information is absolutely required. That problem will be emphasized and explained in more details (and for general averaging operators) in the course of the paper. From the above outline, it can be concluded that the macroscopic path is not well suited for our present objectives. On the other hand, we have also seen that the direct simulation is not tractable. In the language of statistical physics, this direct simulation is a microscopic description since all degrees of freedom are explicitly tracked. A reasonable solution is therefore to choose what can be referred to as a mesoscopic approach, or as a middle-road approach between the microscopic and macroscopic descriptions. The mesoscopic approach retained in this work is a probabilistic approach. Its aim is to model and to simulate the probability density functions of the variables which are of Arst interest. For this reason, the present approach can be deAned as a pdf approach to turbulent dispersed two-phase ows. The di8erent pdfs that will be manipulated are modelled pdfs, that is to say the basis of the approach will be to propose probabilistic models to describe the joint pdf of some variables. It will be seen that probabilistic models can be developed either in terms of the pdf or in terms of the trajectories of the stochastic processes involved. In the present work, we will mainly adopt this second point of view and we will be talking about stochastic particles. The stochastic models will be developed directly for the variables attached to these stochastic particles, providing at the same time
10
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
a Monte Carlo evaluation of the pdf. Thus, the approach can also be referred to as a particle stochastic approach. At the moment, multi-point approaches have not been extended to non-homogeneous wallbounded ows whereas one-point pdf models have been put forward. From the third requirement mentioned above, it follows that one-point pdf models will be often considered. Yet, this is not a strict limitation of the present work. Two-point or multi-point pdfs will be discussed and considered at di8erent stages. This will re ect the fourth constraint since multi-point pdfs represent Aner probabilistic descriptions. Our purpose is a general pdf approach and the relation between these di8erent descriptions is contained in the presentation. 1.3.3. Choice of a rigorous presentation pdf models are already well established in the combustion community. A number of reviews are available which discuss the necessary formalism and present modelling state [6 –9]. In particular, Pope’s work [6] has clariAed the one-point pdf formalism and has given the relations between the Lagrangian and the Eulerian pdfs. In these works, the presentation is strictly tailored for one-point pdf descriptions and the derivation of the stochastic models is often based on the previous knowledge of a macroscopic closure relation [10]. This is probably a reasonable choice (and perhaps the best compromise between model complexity versus tractability) since closures of the reactive source terms, which are local source terms, require only one-point pdfs. In most presentations, the stochastic models are not derived from statistical arguments but from their correspondence with given mean moment equations, although recent proposals have tried to use only arguments from statistical physics [11]. This approach can indeed be regarded as a satisfactory answer for two of the main themes, turbulence and combustion, but application to dispersed turbulent two-phase ows requires further work. Indeed, di8erent physical e8ects are present when discrete particles are considered. Furthermore, the mean equations are not known in advance and should precisely be derived from a probabilistic approach. On the other hand, a particle approach and stochastic models have been used for some time to simulate dispersed two-phase ows, see among others the review of Stock [12] and the references inside. A wide variety of stochastic models have been devised, most of the time from a heuristic point of view. In two-phase ows, the notion of a stochastic particle is, at Arst sight, less surprising than in single-phase turbulence, and it is tempting to skip the careful construction of rigorous foundations since the stochastic concepts may be believed to be ‘evident’. However, this direct approach to stochastic modelling can create severe problems that will be detailed in Section 7. Given that no macroscopic relations are known in advance (and can thus be used as safeguards), the development of a rigorous approach to the pdf description of dispersed turbulent two-phase ows is needed. There is also a new element compared to single-phase reactive ows where the choice of the variables which are explicitly modelled is rather obvious. In two-phase ows, the selection of the basic variables is less obvious and is subject to debate. Then, various choices can be made for a pdf description, and the technical aspects of the hierarchies between these di8erent pdfs must be well understood. As a consequence of the above analysis, the aim of this paper is to build a rigorous probabilistic framework that extends current models developed for single-phase reactive ows to include dispersed turbulent two-phase ows. Such a construction requires a mathematical-oriented
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
11
presentation and the deAnition of a clear methodology. As such, it will be somewhat di8erent from the above-cited references. The presentation will be based on ideas from statistical physics and on the hierarchies of pdf equations with a particle stochastic point of view. The stochastic models will be developed as much as possible from the arguments of statistical mechanics and statistical physics. The purpose of the present work is basically to propose a probabilistic description of a mixed system, composed of a continuous Aeld and of discrete particles. The central notion that is adopted is the Lagrangian point of view which will appear as the ‘propagator of information’ for our complex system. 1.4. Description of the contents The paper has been organized to answer to the following general questions: • • • •
What are the mathematical tools which are required? What are their main characteristics? How are they used for physical modelling in general? How are they precisely used in our case? What do we obtain from them? Are present models the end of the story or can they be improved or coupled to other methods?
These questions correspond to three categories: the mathematics of stochastic processes, the general physical meaning of stochastic modelling, and the development of a speciAc framework. As a consequence, the paper has a three-fold objective. The Arst objective is to provide the reader with a comprehensive and understandable picture of the theoretical tools used in the pdf approach (Section 2). Several notions must be understood: the mathematical properties of stochastic processes, the notion of the trajectories of a stochastic process as well as the correspondence between the trajectory point of view and pdf equations. Once the notion of a Markovian stochastic process (and more precisely the subclass ‘di8usion process’) and its associated pdf is clear, the second objective is the description of the use of di8usion stochastic processes for physical modelling. This is carried out by Arst recalling the concepts of statistical physics, i.e. the N -body problem. A general framework is given to work out the relations between the di8erent levels of contraction (Section 3). Then, the modelling principles that allow stochastic processes to be used are presented (Section 4). From this results the deAnition of a pdf description, which is made up by the choice of the variables which constitute the state vector and by the choice of a stochastic model for this state vector. The third objective concerns the development of a consistent and self-contained framework for the probabilistic description of two-phase ows. This derivation is the core subject of the present work and is performed in four steps. A gradual construction of the complete description has several advantages. It avoids dealing immediately with a complicated formalism which may hide or blur some physical points. By gradually building the complete description, we can discuss at length the physical meaning of the di8erent stochastic terms, for the continuous phase and for the dispersed phase. Since the discrete particles are embedded in a turbulent uid, their motion (and the associated statistical properties) are governed by the underlying turbulent ow. It is then important to detail the physical characteristics of turbulent ows. This is the Arst step of our modelling approach where the reader is given a comprehensive, but still general, overview of the physics of turbulence,
12
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Section 5. The second modelling step is the probabilistic description of single-phase ows, that is the probabilistic description of a continuous Aeld (Section 6). Emphasis is put on the level of information which is needed for successful closures (Kolmogorov theory), on the Lagrangian point of view (which is the natural choice in uid mechanics) and on the existence of a propagator. Correspondences with the Eulerian description and classical mean Aeld equations are given. The third modelling step addresses the question of the probabilistic modelling of discrete particles. The usual issues of particle-tracking models are discussed at length, Section 7. In both the second and the third modelling step, numerical computations are presented and discussed to illustrate how the method performs in di8erent ows and the kind of information that can be extracted from it. The fourth modelling step, i.e. the complete uid–particle pdf approach, is achieved in Section 8. It is shown, once again, that the Lagrangian point of view is the natural choice and that there exists a propagator. Correspondence with Eulerian tools is put forward and mean Aeld equations are derived using rigorous probabilistic arguments. By the end of Section 8, the reader has a clear picture of the pdf approach to turbulent dispersed two-phase ows. Then, the concepts of the probabilistic approach can be summarized and prospects for new developments can be put forward, Section 9. 2. Mathematical background on stochastic processes The purpose of this chapter is to provide clear deAnitions of a stochastic process and of stochastic di8erential equations. These equations appear rather naturally in physical or engineering sciences where one would like to introduce ‘randomness’ or ‘noise’ into the di8erential equations that describe the evolution of a physical system. For example, one would like to give a precise meaning to the equation d Xt = A(t; Xt ) d t + B(t; Xt )t ; dt
(3)
where t is the so-called ‘white-noise’ process that represents some ‘rapid uctuations’. It turns out that the deAnition and proper treatment of such an equation cannot be made directly with classical methods from ordinary di8erential equations (ODEs). Special mathematical notions have to be introduced to explain stochastic calculus which has its subtleties that can be surprising at Arst sight. The following results and notions will be presented, as much as possible, in a logical way while trying to avoid being too mathematically involved. Most of these results will be stated without proofs and not all deAnitions are given. However, complements and detailed presentations of this material can be found in mathematical textbooks [13,14] or in physically-oriented books [15]. An excellent presentation gathering mathematical correctness and an applicationR oriented discussion can be found in Ottinger [16]. Most of the material needed to handle in a simple way probability concepts has been developed in Pope’s seminal work [6] for single-phase pdf methods. It did not appear useful to repeat this presentation here and the objective of this section is to go into more mathematically advanced details. Each of the following subsections cannot pretend to give a comprehensive description of the subjects but the themes and the order of the presentation re ect the important issues.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
13
2.1. Random variables In applied physics, random variables are often introduced directly through their probability density functions (pdf) which can be either discrete or continuous. The random variable, say X , can take a range of possible values x ∈ A, for example x ∈ R or Rd . The probability that X takes a value between x and x + d x is P[x 6 X 6 x + d x] = p(x) d x :
(4)
The actual and more rigorous construction of a random variable is based on an underlying probability space (; F; P) and on measure theory. One deAnes a reference space equipped with a -algebra F (an ensemble of subsets such that complements and reunions of them still give a subset that belongs to the ensemble) on which a measure P (with P() = 1) is deAned. A random variable X is mathematically deAned as a measurable function from this reference space to the one where X takes its values, here A which is also equipped with a -algebra G (; F; P) → (A; G) ; ! → X (!)
(5)
and the law of probability of X is simply the image of the reference measure P, that is PX (A) = P(X −1 (A));
∀A ∈ G :
2.1.1. Conditional expectations The Arst central notion is the expectation of a random variable which is the integral of the possible values against their measures X = X (!) d P(!) : (6) The expected or mean value is written here as X following the usual notation in applied physics but is written as E(X ) in the mathematical literature. The level of abstraction used in the deAnition of random variables is not just for the sake of doing mathematics but is helpful to precise some notions concerning Arst random variables and then for stochastic processes. One such notion is the conditional expectation which is very important for the physical idea of coarse-grained descriptions but can only be fully understood with reference to -algebras. It is worth giving the formal deAnition: Denition 1. If X is a random variable on the probability space (; F; P) and if F is a sub--algebra of F, that is F ⊂ F, the conditional expectation (or conditional average) of X given F , written X |F , is a random variable deAned on the sub--algebra F and such that its expectation or its mean value on any subset A of the sub--algebra F is equal to the mean value of X on the same subset, or X XA = X |F XA
∀A ∈ F ;
where XA is the indicator of the subset A .
(7)
14
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
This formal statement can be translated into words. The Arst important point is that the conditional average is also a random variable but deAned on a coarser -algebra. A -algebra can be regarded as describing mathematically the notion of the ‘level of information resolved’ or the ‘information content’ of the random variable X . If more information is provided on a physical object which is represented by the random variable X , this corresponds in the mathematical deAnition to a function deAned on a ?ner ensemble F. On the reverse, if less information is known or resolved by X this translates into the fact that the function X is deAned on a coarser ensemble F. Therefore, for a physical object represented by a random variable X , the idea of a coarse-grained description, when not enough details are resolved or when one voluntarily disregards some pieces of information, can be well represented by a conditional average. The second important point in the deAnition given above, is that the conditional average is the mean of the actual random variable X ‘averaged’ over the unresolved information or, in more mathematical terms, over the Aner decomposition of any subset A of F into reunions of subsets of F. This appears as the only way to properly deAne the notion of the conditional average of one random variable. However, when one handles in fact two joint random variables X and Y and simply considers the sub--algebras obtained by Axing the value of one of the two random variables, say for example the sub--algebra obtained with Y = y, we retrieve the usual and more intuitive notion of conditional random variable given the value of another one whose pdf is then p(x; y) p(X = x|Y = y) = : (8) pY (y) 2.1.2. Weak and strong convergence of random variables Random variables are not often known directly and are generally obtained as limits of approximate and simpler random variables. This is the case when a process is simulated by numerical integration with Anite time steps Qt, see Section 2.10. This also happens from a physical point of view since models are used to get practical but then approximate answers. One must be able to know how properties or characteristics of the various approximations can be carried over to the actual solution. Several modes of convergence can be deAned for random variables, which must be well understood in particular the distinction between strong and weak convergences. For these reasons, we explicitly give the following deAnitions. Denition 2. The sequence (Xn ) converges towards the random variable X , deAned on the same probability space, almost surely if and only if P({! for which |Xn (!) − X (!)| → 0 as n → ∞} = 1 :
(9)
Denition 3. The sequence (Xn ) converges towards the random variable X , deAned on the same probability space, in the mean square sense if and only if |Xn − X |2 → 0
as n → ∞ :
(10)
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
15
Denition 4. The sequence (Xn ) converges towards the random variable X , not necessarily on the same probability space, in distribution or in law if and only if f(Xn ) → f(X )
as n → ∞ :
(11)
There is actually a fourth possible mode of convergence, the notion of stochastic limit, which is not given here. This mode of convergence is important in the theory but since it will not be explicitly used here and since its absence does not prevent the key concepts from being presented, it is left out so as to limit the mathematical burden. The Arst mode of convergence, the almost sure limit, is the strongest possible one. It is actually the idea of classical pointwise convergence of real-valued functions used for the realizations of the random variable. It means that the sequence (Xn ) converges to X ‘everywhere’, or that the subset on which (Xn ) does not converge to X is negligible (in the sense that the measure of its importance is zero). The second mode of convergence has a more familiar connotation since it manipulates something that is basically an energy. Yet, these Arst two modes are similar. The third one is somewhat di8erent since what is required is that only mean quantities derived from the sequence (Xn ) converge to a mean value derived from the limit process X . That limit process does not need to be known explicitly, and we only deal here with some information extracted from the di8erent processes. That mode of convergence is therefore weaker than the Arst two. Indeed, the Arst two modes depend ‘directly’ on the values of the variables Xn and X whereas in the third mode the knowledge of X is ‘indirect’. Loosely speaking, we can give the overall picture and say, that the almost sure and the mean square convergence are strong modes of convergence while convergence in distribution refers to a weak convergence. The distinction between these two ways of approximating random variables is important within the mathematical theory (deAnition of the Itˆo integral, solutions of equations, : : :) but also for numerical reasons (see Sections 2.10) and for physical purposes since it helps clarifying the ideas of the pdf approach in single- and two-phase ows developed in Sections 6 –8. 2.2. Stochastic processes Another interest of the exact deAnition of random variables given previously is to pave the way for the notion of stochastic processes and of trajectories of a process. A stochastic process is simply a family of random variables X = (Xt ) indexed by a parameter which is usually the time t. This notion is obvious to introduce when one wishes to use random variables to model a time-dependent physical system. The mathematical deAnition of a stochastic function is in fact a measurable function of two variables T × (; F; P) → (A; G) ; (t; !) → X (t; !) :
(12)
The equivalence mentioned in the introduction between di8erent points of view can now be made clear by Axing one of the two variables. (a) for each Axed t ∈ T , Xt is a random variable and we can deAne its pdf p(t; x),
16
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
(b) for each Axed !, we have simply a function t → Xt (!)=X (t; !) which is called a trajectory of the stochastic process X or a sample path, (c) there is a third point of view which generalizes the trajectory point of view. In this point of view, the stochastic process (Xt ) is regarded as a random variable for which the range of values is the set of real functions X: (!). This deAnes the path-integral point of view. In (a), we address the problem by considering the time-dependent pdf and the question is, for a given problem, to write the equation satisAed by this pdf. This is the pdf point of view. In (b), we Arst discretize the reference space introducing ‘particles’ and we follow the time evolution of these particles which deAne the trajectories of the process. This is the trajectory point of view. It is now clear that these particles represent actually di8erent realizations of a stochastic process whose evolution is tracked in time. The path-integral point of view will not be used in the discussion of present models for turbulent dispersed two-phase ows, but will be referred to as an attractive tool in Section 9.3.4. 2.3. Markov processes Manipulation of general stochastic processes is di@cult since it requires to handle N -point distributions, that is the joint distribution functions of the values of the process p(t1 ; x1 ; t2 ; x2 ; : : : ; tN ; xN ) ;
(13)
at N di8erent times, and for any value of N . An important simpliAcation can be obtained for a class of special processes to which we nearly always limit ourselves, Markov processes. A Markov process is deAned as a process for which knowledge of the present is su@cient to predict the future. This is actually a simple notion which is carried over from ordinary di8erential equations (ODE). In classical mechanics, when an ODE is written to describe the time evolution of a system, knowledge of the initial condition is su@cient. For stochastic processes, the Markov property means that if the state of the system is known at time t0 , additional information on the system at previous times s (s 6 t0 ) has no e8ect on the future at t ¿ t0 . The Markov property simpliAes the situation since it can be shown that Markov processes are completely determined by their initial distribution p(t0 ; x0 ) and their transitional pdf p(t; x|t0 ; x0 ). This transitional pdf represents the probability that X takes a value x at time t conditioned on the fact that at time t0 its value was x0 . The Markov property manifests itself in the following consistency relation which is the Chapman–Kolmogorov formula (14) p(t; x|t0 ; x0 ) = p(t; x|t1 ; x1 ) p(t1 ; x1 |t0 ; x0 ) d x1 : This equation states that the probability to go from (t0 ; x0 ) to (t; x) is the sum over all intermediate locations x1 at an intermediate time t1 . The factorization inside the integral re ects the independence of the past and the future at t1 when the present is known. A Markov process can be characterized directly in terms of its transitional pdf or its trajectories or more indirectly (in a weak or distribution sense) by its action on members of a function space. It is useful to deAne the inAnitesimal operator for functions g acting on the sample space
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
17
of Xt by (g(Xt+dt )|Xt = x) − g(x) ; dt→0 dt
Lt g(x) = lim
(15)
where g(X ) denotes the mean or expectation g(X ) = g(x)p(x) d x :
(16)
The value of Lt g(x) can be thought of as the mean inAnitesimal rate of change of the process g(Xs ) conditioned on Xt = x. Using this operator, the Chapman–Kolmogorov relation can be turned into di8erential equations. Since the conditional pdf p(t; x|t0 ; x0 ) is a function of two variables, on can consider variations with respect to the initial variables (t0 ; x0 ) or the Anal variables (t; x). We obtain then two di8erent equations (see [13]), whose meaning will become clearer for di8usion processes: • Kolmogorov backward equation: 9p + Lt p = 0 ; 0 9 t0 end condition p(t; x|t ; x ) = (x − x ) 0 0 0 • Kolmogorov forward equation: 9p = L∗ p ; t 9t initial condition p(t; x|t ; x ) = (x − x ) 0 0 0
(17) when t0 → t :
(18) when t → t0 ;
where Lt∗ denotes the adjoint of the operator Lt . The forward Kolmogorov equation gives the well-known Fokker–Planck equation for di8usion processes as we will see below. 2.4. Key Markov processes 2.4.1. The Poisson process Many situations, such as electron emission, telephone calls, shot noise or collisions, among other problems, require the notion of random points and eventually of Poisson processes. The important properties of the statistics of random points are Arst outlined. If a large number of points n are placed at random within a wide interval, say [ − T=2; T=2], it can be shown that the probability to have k points in an interval I of length tI , small with respect to T , is given by (ntI =T )k : (19) k! We then consider the case when n; T → ∞ such that n=T = # remains Anite. This deAnes the concept of random Poisson points for which the probability to have k points in any interval I of length tI , say n(I ) = k, is thus P(k in I ) = e−ntI =T
P(k in I ) = e−#tI
(#tI )k : k!
(20)
18
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
A very important property is that random points in non-overlapping intervals are independent. The parameter # which speciAes Poisson points has a clear and simple meaning. This is shown by considering a small time interval of length Qt and the probability to have one point within that interval. If Qt is such that #Qt is much less than one, we have P(one point in [t; t + Qt]) #Qt ;
(21)
while we have P(more than one point in [t; t + Qt])Qt :
(22)
Consequently, the parameter # appears as the density of Poisson points. From the concept of random points, the deAnition of the Poisson process is straightforward. The Poisson process, Nt is deAned as the number of random points, or more generally of random events that take place in the interval [0; t] Nt = n(0; t) :
(23)
The Poisson process is therefore a stochastic process taking discrete values. The trajectories of the Poisson process are staircase functions, being constant between random points at which they jump to the next integer. The mean value of the Poisson process and its variance are given by Nt = #t ;
(24)
Nt = (Nt − Nt )2 1=2 = (#t)1=2 :
(25)
The parameter # which deAnes the process is still the density of the random points, or rather of the random times at which certain events take place (emission of an electron, arrival of a phone call, collision with another particle, : : :). It is called the intensity of the Poisson process and has the dimension of a frequency or the inverse of a time scale, say $c . The mean time interval between each random event is simply equal to $c . 2.4.2. Wiener process and Brownian motion The Wiener process is the key process for our present concerns. It represents directly a model for a Brownian particle and as such has direct physical applications for modelling issues. It is also the fundamental building block on which di8usion processes and stochastic di8erential equations are built. The Wiener process can be introduced di8erently, directly through its construction as a random walk in some applied textbooks or as a rather abstract mathematical object in more formal mathematical works. A middle path is sought here and further explanations can be found in [13–15]. We Arst limit ourselves to the one-dimensional case but all results are easily extended to the multi-dimensional case. The Wiener process Wt can be deAned as a Gaussian process. Just as every Gaussian random variable is completely deAned by its mean and variance, a Gaussian process is fully characterized by two functions, its mean and covariance, which are functions of one and two variables respectively: M (t) = Xt ;
(26)
C(t; t ) = (Xt − Xt )(Xt − Xt ) :
(27)
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
19
For the Wiener process, these deAning functions are M (t) = 0;
C(t; t ) = min(t; t )
(28)
and the transitional pdf has a Gaussian form (x − x0 )2 1 p(t; x|t0 ; x0 ) = exp − : 2(t − t0 ) 2((t − t0 )
(29)
The inAnitesimal operator associated to the Wiener process is 1 92 g(x) Lt g(x) = ; (30) 2 9x 2 and the forward Kolmogorov equation shows that the transitional pdf p(t; x|t0 ; x0 ) is the solution of the heat equation 2 9p = 1 9 p(x) ; 9t 2 9x 2 (31) initial conditions p(t; x|t0 ; x0 ) = (x − x0 ) when t → t0 : This equation already reveals the physics of the problem. A quantity will di8use in space (its value follows a di8usion equation such as the heat equation) because it is ‘carried’ by underlying and fast Brownian particles. In other words, the result of the mixing of fast Brownian particles which carry a piece of information is that the mean value of that information di8uses in space. The Wiener process has a number of particular properties. The main ones are: • the trajectories of Wt are continuous yet nowhere di8erentiable. Even on a small interval, Wt
uctuates enormously.
• the increments of Wt , d Wt =Wt+dt − Wt , over small time steps d t are stationary and independent.
Each increment is a Gaussian variable with mean, variance and higher moments given by d Wt = 0;
(d Wt )2 = d t;
(d Wt )n = o(d t) :
(32)
• the Wiener process is the only stochastic process with independent Gaussian increments and
with continuous trajectories. • the trajectories are of unbounded variation in every Anite interval. This property explains why stochastic integrals will di8er from classical Riemann–Stieltjes ones. 2.5. General Chapman–Kolmogorov equations
Some of the typical properties observed with the key stochastic processes described above can be generalized to a whole class of Markov processes, provided that certain assumptions are made on their behaviour over small time increments. From the correspondence between the trajectory and the pdf points of view, there are two ways to express this incremental behaviour. In this section, we follow the trajectory point of view and characterize these processes by the following conditions on the transitional pdf over small increments in time Qt: 1 (33a) p(t + Qt; y|t; x) = W (y|t; x) + O(Qt); for |x − y| ¿ + ; Qt
20
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
1 Qt 1 Qt
|y−x|¡+
|y−x|¡+
(y − x) p(t + Qt; y|t; x) d y = A(t; x) + O(Qt) ;
(33b)
(y − x)2 p(t + Qt; y|t; x) d y = B2 (t; x) + O(Qt) :
(33c)
The Arst condition is the probability of a jump and trajectories are discontinuous when W = 0. The second one deAnes the drift coe@cient A(t; x) which is the mean increment of the conditional process Xt . The third one deAnes the di8usion coe@cient which represents the variance of the increment or the spread around the mean incremental value. The transitional probability density function p(t; x|t0 ; x0 ) is a function of both the initial state (t0 ; x0 ) and of the Anal state (t; x). Consequently, two points of view can be adopted by holding either the initial or the Anal condition Axed and by varying the other state. Using the Chapman–Kolmogorov relation, Eq. (14), and the above hypotheses, it can be shown that, when the initial condition is held Axed and when p is regarded as a function of the Anal state (t; x), then p(t; x|t0 ; x0 ) satisAes the forward Kolmogorov equation [15] 9p 9[A(t; x) p] 1 92 [B2 (t; x) p] =− + 9t 9x 2 9x 2 + {W (x|t; y) p(t; y|t0 ; x0 ) − W (y|t; x) p(t; x|t0 ; x0 )} d y :
(34)
Using similar considerations and more or less the same derivation, it can be shown that, as a function of the initial state (t0 ; x0 ) when the Anal condition (t; x) is held Axed, p(t; x|t0 ; x0 ) satisAes the backward Kolmogorov equation [15] 9p(t; x|t0 ; x0 ) 9p(t; x|t0 ; x0 ) 1 2 92 p(t; x|t0 ; x0 ) = −A(t0 ; x0 ) − B (t0 ; x0 ) 9t0 9x0 2 9x02 + W (y|t0 ; x0 ){ p(t; x|t0 ; x0 ) − p(t; x|t0 ; y)} d y :
(35)
It is important not to confuse the two points of view (forward or backward) which further justiAes the central role of the transitional pdf. In the forward equation, the initial state does not appear explicitly in the jump, drift and di8usion coe@cients, and we can integrate over all possible initial conditions. Since the pdf of the stochastic process Xt at time t is of course given by p(t; x) = p(t; x|t0 ; x0 )p(t0 ; x0 ) d x0 ; (36) it follows that p(t; x) satisAes the same forward equation. From the general Chapman–Kolmogorov equations, various cases can be isolated by considering di8erent possibilities for the jump, drift and di8usion coe@cients. These particular equations have sometimes been obtained separately and carry di8erent names often for historical reasons. Yet, in the present formulation, they appear as subclasses of a general class of Markov processes.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
21
2.5.1. The Master equation When A(t; x) = B(t; x) = 0, the stochastic process involves only jumps and between each jump the trajectories of the process are straight lines. The pdf equation is called the Master equation 9p (37) = [W (x|t; y)p(t; y|t0 ; x0 ) − W (y|t; x)p(t; x|t0 ; x0 )] d y : 9t The Master equation is indeed the central equation for processes which are typically discrete and is met when one deals with physical issues which are also by nature discrete. The classical example is molecular and chemical processes which involve either a complete change or no change at all. Another typical application is for particle collisions where particle velocities can be constant and change discontinuously at discrete and random times. Over a Anite time step Qt, we have p(t + Qt; y|t; x) = (y − x) 1 − W (y |t; z) d y + W (y|t; x)Qt (38) which shows that W (y|t; x) is the probability to jump from state x to y at time t per unit of time. The generic process in this subclass is the Poisson process described in the previous section for which the sample space is discrete and W (x + 1|t; x) = #. 2.5.2. The Liouville equation When W (y|t; x) = B(t; x) = 0, the process is a continuous deterministic process and the pdf equation is the Liouville equation 9p 9[A(t; x)p] =− : (39) 9t 9x The Liouville equation is central in classical mechanics. Its characteristic form, and the presence of only Arst-order partial derivatives, are closely related to the choice of a closed description of a mechanical system (each degree of freedom is explicitly tracked) as it will be explained in detail later on in Sections 3 and 4. 2.5.3. The Fokker–Planck equation When W (y|t; x) = 0, the forward Kolmogorov equation is called the Fokker–Planck equation. 9p 9[A(t; x)p] 1 92 [B2 (t; x)p] : =− + 9t 9x 2 9x 2
(40)
Compared to the Liouville equation, the Fokker–Planck equation involves a supplementary term with a second-order partial derivative. The existence of this term has deep consequences both mathematically and physically. From the mathematical point of view, the issue is to deAne clearly the corresponding behaviour of the trajectories of the process and to put the manipulation of these trajectories on a sound footing. From the physical point of view, the issue is to explain how this behaviour comes into play and the physical meaning behind its use. That question is addressed in Section 4. The solutions of Fokker–Planck equations are known as di8usion processes and the rest of the present section is devoted to clarifying their characteristics and how they are manipulated. The central example within the subclass of di8usion processes is the Wiener process, described in the previous section.
22
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
2.6. Stochastic diCerential equations and diCusion processes Di8usion processes form a subclass of Markov processes. They have been carefully studied and their properties are rather well-known mathematically which makes them easier and safer to use in applied physics. They will be used extensively for modelling purposes since they represent models for di8usion phenomena (thus their name) and have continuous trajectories. The Arst example is the Wiener process described above and general di8usion processes are in fact extensions of it. The inAnitesimal operator for such a di8usion process is given by 9 92 1 L = A(t; x) + B2 (t; x) 2 (41) 9x 2 9x and the transitional pdf p(t; x|t0 ; x0 ) satisAes the Fokker–Planck equation, Eq. (40), with the initial condition p(t; x|t0 ; x0 ) → (x − x0 ) when t → t0 . The Fokker–Planck equation re ects the pdf point of view. On the other hand, the second point of view will give direct answers to the questions explained in the introduction of this chapter related to the meaning of Eq. (3). One would like to give a meaning to the ‘noise’ term, t , introduced in an ODE. The proper way to do so is to say that we are now dealing with a stochastic process Xt and that we are writing di8erential equations for the trajectories of this process as deAned above. We consider now t as a rapidly uctuating, highly irregular stochastic process. The ideal model is a Gaussian ‘white noise’ model where the process is stationary with zero mean and no correlation, that is t = 0
and
t t = (t − t) :
(42)
This process has a constant spectral density (thus the name white noise). However, the white-noise process cannot be deAned directly since it has an inAnite variance. One can give an abstract sense to this process (Arnold, Chapter 3). However, there is a simpler way out of this di@culty. The solution consists in considering the e8ect of the white-noise term over (small) time intervals. We deAne t Yt = t d t : (43) 0
Yt is a Gaussian Markov process whose mean and covariance functions are worked out from the properties of the white noise Yt = 0
and
(Yt )2 = t :
(44)
Therefore, Yt can be identiAed with the Wiener process, Yt = Wt . This indicates that in fact, the integral over a time interval of the white-noise process gives the Wiener process and this justiAes writing t Wt = t d t or d Wt = t d t : (45) 0
The idea is thus to try to deAne not the derivatives of the trajectories, Eq. (3), but their increments over small time steps as d Xt = A(t; x) d t + B(t; x) d Wt ;
(46)
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
which is a short-hand notation for t t Xt = Xt0 + A(s; Xs ) d s + B(s; Xs ) d Ws : t0
23
(47)
t0
The Arst integral can be thought of as a classical one. The second one is a stochastic integral which must be properly deAned. In the usual sense, one would split the time interval into a number of small time steps and write the integral as the limit t N B(s; Xs ) d Ws = lim B($i ; X$i )(Wti+1 − Wti ) ; (48) t0
N →∞
i=1
where $i is chosen in the interval [ti ; ti+1 ]. However, it turns out that the limit result is not independent of the choice of the intermediate time $i as one would expect from classical integration. Di8erent choices of this intermediate time yield Anite but di8erent results for the integral. This surprising result can be traced back to the ragged behaviour of the Wiener process and to the fact that its trajectories are of inAnite total variation in any time interval. To obtain a meaningful and coherent theory (for later manipulations) of the stochastic integral, one must therefore choose the intermediate times right from the outset. The important message here is that speaking of a stochastic integral without specifying in what sense it is considered is not meaningful. Two main choices have been made in the literature. The Arst one is called the Itˆo deAnition and consists in taking the value at the beginning of the time interval $i = ti . There is a clear probabilistic interpretation of this. The integral writes t N B(s; Xs ) d Ws = lim B(ti ; Xti )(Wti+1 − Wti ) (49) t0
N →∞
i=1
which shows that we consider the function B(t; Xt ) as a non-anticipating function with respect to the Wiener process. The choice of $i signiAes that we express B(t; x) as a function of the present state while the increment d Wt which is independent of the present is said to ‘point towards the future’. This choice is in fact rather natural when the ‘noise’ does not depend on the system. From it, result the properties of the Itˆo stochastic integral
t1 Xt d Wt = 0 ; (50) t0
t1 t3 t1 for t0 6 t2 6 t1 6 t3 Xt d Wt Yt d Wt = Xt Yt d t : (51) t0
t2
t2
The second choice is to take the intermediate point $i as the middle point of the interval $i = (ti + ti+1 )=2. This results in the Stratonovich deAnition. Actually, various deAnitions of the Stratonovich integral can be found depending upon the exact expression of the term involving B($i ; X$i ) in the limit sum, Eq. (48). For example, one can choose to take B((ti + ti+1 )=2; X(ti +ti+1 )=2 ) or B(ti ; X(ti +ti+1 )=2 ) as in Arnold [13]. The most common deAnition met in mathematical books is (written with a characteristic symbol ◦) t N 1 B(s; Xs ) ◦ d Ws = lim (52) [B(ti ; Xti ) + B(ti+1 ; Xti+1 )](Wti+1 − Wti ) : N →∞ 2 t0 i=1
24
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
These various deAnitions di8er only by the assumptions required on B for the sums to converge, and if B is smooth enough they lead to the same limit object. Therefore, the above sum can be taken as the present deAnition of the Stratonovich integral. The question of what deAnition of the stochastic integral, Itˆo or Stratonovich, should be chosen has led to continuous debate in applied physics textbooks. A detailed discussion on this dilemma is outside the scope of the present notes. However, the key point is to be aware of this apparent peculiarity so as to avoid confusion. Indeed, if these di8erent deAnitions and properties are ignored, it is hard to understand why calculations performed with seemingly identical procedures can lead to con icting results. Actually, the two forms can be transformed one into the other. The Stratonovich deAnition of a SDE d Xt = A(t; x) d t + B(t; x) ◦ d Wt ;
(53)
can be shown to be equivalent to the Itˆo SDE [13,15] 9B(t; x) 1 d Xt = A(t; x) d t + B(t; x) d t + B(t; x) d Wt : (54) 2 9x The di8erence between the two deAnitions is therefore a mean drift term and is not ‘negligible’. This illustrates and further stresses that, even if one is not interested in mathematical subtleties, a careful deAnition and at least some understanding of what these deAnitions embody is unavoidable. The best illustration of such pitfalls is perhaps numerical schemes for the integration of the trajectories of the process in practical computations, see Section 2.10. Finally, for a stochastic process Xt whose trajectories satisfy stochastic di8erential equations in the Stratonovich sense, it can be seen from the correspondence with an Itˆo form, Eq. (54), and the Fokker–Planck equation veriAed for di8usion processes in the Itˆo sense, Eq. (40), that the pdf of Xt is the solution of
9p 9[B(t; x)p] 9[A(t; x)p] 1 9 B(t; x) : (55) =− + 9t 9x 2 9x 9x 2.7. Stochastic calculus Most of the strangeness of stochastic processes and of SDEs is embodied in stochastic calculus. Although surprising at Arst sight, the di8erences with ordinary di8erential rules are not too di@cult to grasp. They stem from the irregular behaviour of the trajectories of the Wiener process Wt . Indeed, we have seen that on a small time increment d t the variance of the increments of the Wiener process, (d Wt )2 , is linear in d t (in fact, it is equal to d t). This is already contradictory with the ‘normal’ calculus result which says that the square of an increment should be of order (d t)2 . The explanation is that the ‘correct’ behaviour is expected for a di8erentiable process (a process whose trajectories are di8erentiable) while the Wiener process is precisely not di8erentiable. As a consequence, normal calculus rules must be modiAed by going over to the second-order derivatives, which in normal cases give only terms of order (d t)2 but will bring a Arst-order contribution in our case. To illustrate this, we consider a SDE deAned in the Itˆo sense d Xt = A(t; x) d t + B(t; x) d Wt
(56)
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
25
and we want to derive the SDE veriAed by a function g(t; Xt ) of the stochastic process Xt . The rule of thumb is thus to write the Taylor series up to the second order and not to forget the contribution that arises from the term involving (d Wt )2 whose mean is d t. The result is the Itˆo’s formula 9g 9g 9g d g(t; Xt ) = (t; Xt ) d t + A(t; Xt ) (t; Xt ) d t + B(t; Xt ) (t; Xt ) d Wt 9t 9x 9x 92 g 1 + B2 (t; Xt ) 2 (t; Xt ) d t ; (57) 2 9x where the last term on the second line is the ‘new term’ with respect to classical calculus which would have produced only the Arst line. On the other hand, the choice of the Stratonovich deAnition leads to calculus rules which are identical to classical ones [13,15]. However, this nice point is o8set by the di@culty in manipulating the stochastic integral and Itˆo’s simple properties Eqs. (50) – (51) are no longer valid.
2.8. Langevin and Fokker–Planck equations Once stochastic calculus has been deAned and the signiAcation of SDEs has been given, the picture is complete. We can state what is in fact the main point of this whole section: when dealing with stochastic processes there are two ways to characterize the properties, the time-evolution equation of the trajectories of the process or the equation satisAed in sample space by its pdf. This correspondence is particularly clear for di8usion processes and is central in the present paper. We use this summary as an opportunity to write results in the multi-dimensional case. If Z(t) = (Z1 ; : : : ; Zn ) is a di8usion process with a vector drift A = (Ai ) and a di8usion matrix B = Bij , the trajectories of the process are solutions of the following SDE d Zi = Ai (t; Z(t)) d t + Bij (t; Z(t)) d Wj ;
(58)
where Wt =(W1 ; : : : ; Wn ) is a set of independent Wiener processes. The SDEs are called Langevin equations in the physical literature. This corresponds in sample space to the Fokker–Planck equation for the transitional pdf written p(t; z|t0 ; z0 ) 9p 9[A(t; z)p] 1 92 [(BBT )ij (t; z)p] + ; (59) =− 9t 9 zi 2 9zi 9zj where BT is the transpose matrix of B. Actually, the correspondence between the two points of view is not a strict equivalence. Indeed, the matrix D that enters the Fokker–Planck equation is related to the di8usion matrix of the SDEs B by D = BBT . Since there is not always a unique decomposition of deAnite positive matrices for a given matrix D, there may exist several choices for the di8usion matrix B. Therefore, we can have di8erent models for the trajectories that still correspond to the same transitional pdf. In other words, there is more information in the trajectories of a di8usion process than in the solution of the Fokker–Planck equation. However, since we are in the present work interested mainly in statistics extracted from the stochastic process, or in a weak
26
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
approach (in the sense already used in Section 2.1.2), we can consider that the di8erent models for the trajectories belong to the same class and then speak of the equivalence between SDEs and Fokker–Planck equations. 2.9. The probabilistic interpretation of PDEs The equivalence between the trajectory and the pdf points of view is also the basis of the probabilistic interpretation of some PDEs. The starting idea is to interpret the solution of a PDE as a function or a functional of some stochastic process. Instead of solving the PDE by classical numerical methods, the idea is then to simulate directly the trajectories of the process and to obtain the solution by some sort of averaging operation. This methodology can be applied to a large variety of PDEs (see [14,17]). We limit ourselves to the case of parabolic equations and the relation between stochastic processes and PDEs of convection–di8usion type is simply the relation between the two deAnitions of a di8usion process. The probabilistic interpretation reverses that point of view and regards a convection–di8usion PDE as a kind of Fokker–Planck equation. For example, the solution of the problem 2 2 9u = − 9[A(t; x) u] + 1 9 [B (t; x) u] ; 9t 9x 2 9x 2 (60) u(0; x) = h(x); when t = 0 ; can be built from the transitional pdf of the di8usion process Xt , p(t; x|t0 ; x0 ), as u(t; x) = p(t; x|t0 ; x0 )h(x0 ) d x0 ;
(61)
where A(t; x) and B(t; x) are, respectively, the drift and di8usion coe@cients of the process Xt . Therefore, in physical terms, Xt appears as the propagator of the initial function h(x0 ). Or in other words, Xt is the carrier of the information. At the initial time, particles start at x0 with an ‘information’ that is h(x0 ). Then they follow the SDE d Xt = A(t; x) d t + B(t; x) d Wt :
(62)
As a consequence of this motion, information is carried from the initial state (t0 ; x0 ) to another one (t; x). The average result is then the solution of the PDE which is of convection–di8usion type. It is seen that the di8usion term in the PDE re ects in fact the fast and random motion expressed by the Wiener process, d Wt , in the ‘particle’ evolution equation. Conversely, when particles undergo a random walk, the result of their mixing is to produce a di8usion in space. Then, in practical simulations, any statistics that are continuously obtained from the pdf of the process, can be approximated, at a given time t, from an ensemble of realizations of the process by the Monte Carlo evaluation N 1 f(Xt ) f(Xti ) : (63) N i=1
For a stochastic process when statistics are required at various times, the di8erent realizations at time t are simply provided by the values at the corresponding time of the trajectories of the
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
27
process. Indeed, it is now clear that the trajectory point of view consists in practice in following in time a number of trajectories, that we can write as X i (t) = X (t; !i ) for a certain number of possible events represented here by !i . In other words, simulating stochastic processes from the trajectory point of view corresponds to performing Monte Carlo integration at each time. One speaks then of the Monte Carlo integration of partial di8erential equations. 2.10. A word on numerical schemes The problem of how to devise accurate numerical schemes for the integration of SDEs is a di@cult issue, and also a recent concern. This is the subject of current research [18,19]. The detailed presentation of current state-of-the-art proposals is not within the scope of the present paper and we limit ourselves to the main points that also illustrate the notions put forward in the previous sections. Compared to similar numerical schemes that are now well established for ordinary di8erential equations, the question of the consistency of stochastic numerical schemes must be carefully analysed. Actually, most of the di@culties arise from a lack of understanding of the exact deAnition of the stochastic integral, see Section 2.7. Numerical schemes, as well as manipulation of a function of the stochastic process Xt can only be done after an interpretation of the stochastic integral has been chosen. If one has chosen the Itˆo interpretation, then it is implicitly assumed that the discretization of B(t; x) should not anticipate the future. As a result, Runge–Kutta schemes cannot be applied directly. More precisely, careless applications of high-order Runge– Kutta schemes can introduce spurious drifts which may not be easy to detect. For the Langevin equation d Xt = A(t; Xt ) d t + B(t; Xt ) d Wt ;
(64)
the Euler scheme is the simplest choice and is written as X i (t + Qt) − X i (t) = A(t; X i (t))Qt + B(t; X i (t))QWt ; (65) √ where the random term QWt is expressed as Qt × e, e being a value sampled in a normalized Gaussian random variable, independently at each time step and for each trajectory. A rather illuminating example of typical pitfalls is seen if one tries to apply directly the well-known predictor–corrector scheme. This is a two-step scheme with the Euler scheme acting as a predictor i X˜ (t + Qt) − X i (t) = A(t; X i (t))Qt + B(t; X i (t))QWt ;
(66a)
1 i X i (t + Qt) − X i (t) = (A(t; X i (t)) + A(t + Qt; X˜ (t + Qt)))Qt 2 1 i × (B(t; X i (t)) + B(t + Qt; X˜ (t + Qt)))QWt : (66b) 2 Yet, a time series expansion of this scheme reveals that due to the Arst-order behaviour of (QWt )2 in time, the corresponding di8erential equation turns out to be Eq. (54) rather than the Itˆo SDE which is here Eq. (64). In other words, the predictor–corrector scheme is consistent, however with the Stratonovich interpretation of SDEs, Eq. (53), but not (in general) with the
28
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Itˆo interpretation. Therefore, if the Itˆo interpretation has been chosen and the stochastic integrals are manipulated using the simple Itˆo’s rules (see Section 2.6), the scheme is not consistent. The key point here is that the numerical discretization must be in line with the mathematical deAnition of the stochastic terms. To stay on somewhat safer grounds, one can stick to the Euler scheme or pay enough attention to the validity of the numerical schemes. After consistency is checked, the quality of schemes must be measured to analyse how they actually approximate solutions, and for this the notion of order of convergence must be properly deAned. For stochastic processes various deAnitions can be adopted which mirror the di8erent ways random variables may converge to a limit random variable, see Section 2.1.2. One can deAne a strong order of convergence and a weak order of convergence. Let us consider a numerical approximation of the process Xt obtained with a Anite time step Qt, called XtQt . On the one hand, the numerical scheme will have a strong order of convergence m if at a time tmax we have that |Xtmax − XtQt |2 1=2 6 C(Qt)m : max
(67)
On the other hand, the numerical scheme will have the weak order of convergence m if at time tmax we have that |f(Xtmax ) − f(XtQt )| 6 C(Qt)m max
(68)
for all su@ciently smooth functions f. For example, the Euler scheme has a strong order of convergence m = 1=2 but a weak order of convergence m = 1. As already explained before, since we are mainly interested in approximating various statistics of single- and two-phase ows the natural notion is the notion of weak convergence. 3. Hierarchy of pdf descriptions Most of the necessary mathematical elements concerning stochastic di8erential equations have been given in the preceding section. For our purposes, attention has been focused on Markovian processes, and more speciAcally on a particular subset of Markovian processes, di8usion processes. These processes will be used as building blocks, Arst in turbulent single-phase ow modelling in Section 6 and then in turbulent two-phase ow modelling in Sections 7 and 8. Up to now, emphasis has been mainly put on the mathematical characteristics of di8usion processes rather than on their application for physical purposes. Such an application requires further analysis and discussion. Indeed, even in the multi-dimensional case where the stochastic process Z(t) is a vector of d real stochastic processes Z(t)=(Z1 (t); : : : ; Zd (t)), the selection of variables that make up the stochastic process Z(t) in a practical case, its dimension d, and the choice of the evolution equation (through the drift and di8usion coe@cients), were not discussed and were considered as given. However, a pdf description appears in a closed form only when: (i) the stochastic process Z(t) is chosen, (ii) the model for its time evolution equation is speciAed. The form and the nature of the di8erent models used for two-phase ow modelling will be presented in detail in later sections. In the present one, we discuss issues related to the choice of
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
29
the stochastic process Z(t) which is used to describe a physical system. By considering di8erent stochastic descriptions, either Aner or coarser, di8erent pdf equations result. It is important to be aware of the interplay between the di8erent and increasingly coarser descriptions and the structure of the corresponding reduced pdf equations. In practice, the cornerstone of the contracted description is, of course, to be able to choose the ‘correct’ reduced number of variables, which must be small enough to make up a tractable system while still capturing the essence of the physics of the problem. The discussion on how to perform such a choice in some cases is postponed to the next section. In the present one, we limit ourselves to the technical presentation of this interplay which manifests itself by various pdf hierarchies. These hierarchies will be referred to continuously in the rest of the paper. The general issue of a pdf hierarchy is Arst presented, and is then illustrated by two examples. The Arst hierarchy is very well known in Statistical Physics. However, the second hierarchy is not often described, though it is of the same nature. Both hierarchies appear constantly in the modelling considerations later on. 3.1. Complete and reduced pdf equations Numerous physicals situations fall into the category of what is called N -body problems. That is, we have N objects, identical or not, which interact mutually. This situation can be loosely referred to as a N -particle problem by deAning each ‘object’ as being a particle. This terminology will be retained here. In this general approach, each particle represents the particular value of a set of variables and is fully determined by the knowledge of these ‘internal’ variables. A classical example is molecular dynamics problems, where each particle represents a molecule and can be thought of as a point particle deAned by the value of its location and velocity. In another case, the knowledge of the state of each particle may require more variables. The way these N particles interact and in uence one another is considered to be known when the state of the N particles is known, that is the mutual forces are internal with respect to the whole system made up by the ensemble of the N particles. The dimension of the system (or the number of degrees of freedom), d = dim(Z), is given by d = N × p, where N is the number of particles included in the system and p represents the number of variables attached to each particle. For this system, the complete vector which gathers all available information is then Z = (Z11 ; Z21 ; : : : ; Zp1 ; Z12 ; Z22 ; : : : ; Zp2 ; : : : ; Z1N ; Z2N ; : : : ; ZpN ) : This vector is the state vector of the N -particle system. The vector deAned by the p variables attached to each particle, Zi = (Z1i ; Z2i ; : : : ; Zpi ), is called the one-particle state vector, in this case for the particle labelled i. In practice the dimension of the system is huge (it might be inAnite) and one has to come up with a reduced (or contracted) description, or in other words to consider a subset of dimension d = s × p d. Such a reduced description is needed to achieve a practical formulation of the behaviour of the system, that is to formulate a set of equations in closed form which can be solved numerically with help of modern computer technology. The key point is that, in the general case, such a contraction is followed by a loss of information and that knowledge of higher-order pdfs has to be provided through closure relations.
30
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
To illustrate this problem, let us consider a N -particle system where the time evolution equation involves simply a deterministic force d Z(t) = A(t; Z(t)) : (69) dt The dimension of the complete state vector Z is equal to d, and the corresponding pdf p(t; z) veriAes the Liouville equation 9p(t; z) 9 + (A(t; z)p(t; z)) = 0 : (70) 9t 9z This equation is closed since in fact all the degrees of freedom of the system are explicitly tracked. We consider now a reduced pdf pr (t; zr ) where dim(Zr ) = d and p(t; z) = p(t; zr ; y) with, of course, dim(Y) = d − d . By integration of the previous equation on y, the transport equation for the marginal (reduced) pdf becomes 9pr (t; zr ) 9 (71) + r [A|zr pr (t; zr )] = 0 ; 9t 9z where the conditional expectation is deAned by 1 r r r A|z = A(t; z ; y)p(y|t; z ) d y = A(t; zr ; y) p(t; zr ; y) d y : (72) p(t; zr ) Eq. (71) is now unclosed. This illustrates the fact that when a reduced description (in terms of a subset of degrees of freedom) is performed, information is lost, and one has to come up with a closure equation for higher-order pdfs. We have moved from a complete description and therefore a closed pdf equation Eq. (70), to a contracted description and thus an unclosed pdf equation Eq. (71). At this point, two sets of reduced descriptions can be chosen in the N -particle example, by varying either the number of particles retained in the state vector of the reduced system or by varying the number of variables attached to each particle. The Arst one corresponds to the classical BBGKY hierarchy (the initials are those of the authors who derived it independently: Bogoliubov, Born, Green, Kirkwood and Yvon) encountered in kinetic theory (p = 2), and is fully described in textbooks, for example [20,21]. In the second one, the dimension of the state vector is addressed from a single particle point of view, s = 1. 3.2. BBGKY hierarchy Classical mechanical questions are well represented by N -particle deterministic problems, involving N particles of identical mass m in mutual interaction and with no external forces. The dimension of the one-particle state vector is, almost always, taken as two, including particle location and velocity. This is a consequence of the search of a kinetic description and of the hypothesis that forces derive from a location-dependent potential. Consequently, the dimension of the complete state vector is d = 2 × N . The drift vector is A = (U; F) where the mutual acceleration, taken in the direction xi − xj , is denoted Fij and is given in terms of a potential
ij = (|xi − xj |), which is the mutual potential energy of the pair of particles (i; j). Therefore mFij = 9 ij = 9xi represents the force on particle i due to particle j. In the classical mechanical framework, a reduced description is meant as a description of the system using identical variables for each particle but using only a subset of the total number. The reduced pdf for a subset of
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
31
s particles, ps (t; y1 ; V1 ; : : : ; ys ; Vs ) is written for the sake of simplicity as ps (t; 1; : : : ; s) = ps and consequently for integration d ys d Vs reads d s. Integration of the Liouville equation yields (summation over the i index should not be confused with tensor notation as it represents the number of particles in the subset, that is a summation from 1 to s; 1 6 s 6 N ) 9 Ls (ps ) + Fij pN d (s + 1) : : : d N = 0 ; (73) 9Vi j¿s where the Ls operator is given by Ls (·) =
9· 9 9 (Vi ·) + + 9t 9yi 9Vi
s
Fij · :
(74)
j=1
Eq. (73) has been obtained by applying the correspondence di8usion process—Fokker–Planck equation and more especially deterministic process—Liouville equation. It can also be derived using Classical Mechanics, i.e. the properties of the Liouville operator, Libo8 [21], or the Hamiltonian, Balescu [22]. Noticing that (by permutation and variable changes) N Fij p d (s + 1) : : : d N = (N − s) Fi(s+1) ps+1 d (s + 1) ; (75) j¿s
the following set of equations is obtained: 9 Ls (ps ) + (N − s) Fi(s+1) ps+1 d (s + 1) = 0 ; 9Vi
(76)
which is a set of N coupled equations and is often called the BBGKY hierarchy. This simply states that for a deterministic ensemble of N particles, a contracted description of the system gives an unclosed equation on the reduced pdf as illustrated by Eq. (76). For s = 1, one-point pdf, one recognizes the kinetic equation which involves the two-point pdf and so on. At this point, it should be mentioned that, in the case of mutual interactions given by a simple potential, it was quite trivial to illustrate the hierarchy of pdfs but, for example in the case of discrete particles (or even uid particles) carried by a turbulent uid, the expression of the force exerted on a particle does not exhibit a simple analytical form as it depends simultaneously on all other particles and consequently, in this case, the hierarchy problem is given by Eq. (71). At last, this type of hierarchy is not a property of the pdf approach but is typical, in general, for problems where a reduction is made, as for example, in the case of the Reynolds decomposition of the local instantaneous Navier–Stokes equations. 3.2.1. Normalization of the distribution function In the previous approach, a pdf, p(t; x), has been used, p(t; x) d x is in fact the probability to And the system (the N particles) in a given state in the range [x; x + d x], cf. Section 2.2 (this can be understood more easily using the notion of an ensemble density function, introduced by Gibbs, cf. e.g. [21]). The marginal ps represents then the probability to And the reduced system (s particles) in a given state.
32
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
In many applications, as will be seen later, it is convenient to work with the s-tuple distribution function, fs (t; 1; : : : ; s), where fs (t; 1; : : : ; s) d 1 : : : d s represents the probable number of s-tuple in a given state in the range [1; 1 + d 1]; : : : ; [s; s + d s] at time t. The relation between ps and fs is directly given by combinatorics, that is by the number of ways of taking s elements from a population of N elements, without replacement and of course with regard to order. The answer is (N )k , that is N! fs (t; 1; : : : ; s) = (77) ps (t; 1; : : : ; s) : (N − s)! Normalization is given by fs (t; 1; : : : ; s) d 1 : : : d s =
N! (N − s)!
ps (t; 1; : : : ; s) d 1 : : : d s =
N! ; (N − s)!
(78)
and with r ¡ s 6 N , the r- and s-tuple distribution functions verify the following relation: (N − s)! r f (t; 1; : : : ; r) = fs (t; 1; : : : ; s) d (r + 1) : : : d s : (79) (N − r)! The BBGKY hierarchy, Eq. (76), can be written in a slightly di8erent form 9 s s s+1 Fi(s+1) f d (s + 1) = 0 : L (f ) + 9Vi
(80)
3.3. Hierarchy between state vectors The BBGKY hierarchy gives a comprehensive picture of the resulting modelling problem in the frame of Classical Mechanics. The issue is now to express the statistical e8ect of all the disregarded particles on the statistical properties of the small number (usually one or two) of particles that are kept in the state vector. In this hierarchy, the choice of the one-particle state vector and its dimension, here p = 2, remains unchanged. However, in di8erent situations, various choices can be made for the one-particle state vector and it is useful to consider a second set of pdf equations which corresponds to di8erent and increasing one-particle state vectors. This happens already when we consider a N -particle problem where the force acting on one particle due to the other ones can be any function of particle properties, for example a function of particle acceleration or other ‘internal’ particle properties. It is therefore important to express also the interplay between the choice of the one-particle state vector and the structure of the corresponding pdf equation, even when a given subset of s particles is considered. There is another strong justiAcation for considering this second pdf hierarchy with respect to modelling purposes. Indeed, to obtain a closed pdf equation at some chosen level, a model must be introduced to simulate the behaviour of the degrees of freedom that are summed over. As will be explained more in detail in the following section in the case of white-noise terms, it is important to select the ‘correct’ variable that can be well modelled by a certain stochastic process. A very precise example of this choice will be given by the choice of the variable to model in one-point particle pdf for two-phase ows, see Section 7. The BBGKY hierarchy was presented using a top-bottom approach, that is starting from the complete Liouville equation and deriving from it the di8erent reduced descriptions. The
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
33
hierarchy between vector states will be presented here from a bottom-top approach, starting from the most reduced level to higher level and introducing modelling concerns. We consider only one particle (s = 1), and follow a presentation based on the historical case of a Brownian particle that will be taken up again in the next section. First of all, we can restrict ourselves to following the position of the particle (that was Einstein’s point of view with time steps that are large enough, see the next section). With that choice of the state vector Z(t) = (X(t)), the particle velocity is an external variable and the pdf equation for p(t; y) is unclosed 9p(t; y) 9 (81) + (U|yp(t; y)) = 0 : 9t 9y To obtain a closed model, the e8ect of the particle velocity has to be replaced by a model d X + (t) d X (t) = U + (t) ⇒ = F[t; U (t)] dt dt
(82)
where the superscript + denotes the exact equation and F[t; X (t)] represents a functional of the position X (t). If the functional F is deterministic we end up with a reduced Liouville equation. However, if F is stochastic, the techniques of Section 2 may be applied. If this Arst picture is believed to be too crude, one can include the velocity of the particle in the state vector that becomes then Z(t) = (X(t); U(t)) (Langevin’s point of view). Now, the particle acceleration A(t) becomes an external variable and the corresponding pdf equation for p(t; y; V) is unclosed 9p(t; y; V) 9(Vi p(t; y; V)) 9 + (A|y; Vp(t; y; V)) = 0 : (83) + 9t 9yi 9Vi To obtain a closed form, the acceleration has to be eliminated or replaced by a model d X + (t) d X (t) + = U (t) = U (t) dt dt ⇒ + d U (t) = A+ (t) d U (t) = F[t; X (t); U (t)] : dt dt It is thus clear that the second description encompasses the Arst one. It contains more information and in physical terms corresponds to a description performed with a Aner resolution. From a modelling point of view the task is also di8erent depending upon the choice of the one-particle state vector. In the Arst case (Einstein’s point of view), one has to model particle velocities. In the second case (Langevin’s point of view) one has to model particle accelerations. From the above example, a general picture emerges. We consider a one-particle reduced description (s = 1) but with many internal degrees of freedom, i.e. Z1 = (Z11 ; Z21 ; : : : ; Zp1 ; : : :). The complete one-particle state vector is written here for a particle labelled i = 1, but in a one-particle pdf description the label is irrelevant (the same would be valid for any particle i) and the superscript is therefore skipped in the following. If the time rate of change of the particle degrees of freedom has the following form: d Z1 (85a) = g(t; Z1 ; Z2 ) ; dt d Z2 (85b) = g(t; Z1 ; Z2 ; Z3 ) dt
34
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
.. .
(85c)
d Zp = g(t; Z1 ; : : : ; Zp ; Zp+1 ) dt
(85d)
.. .
(85e)
and if the chosen one-particle reduced state vector contains only a limited number of degrees of freedom, say p; Z r = (Z1 ; : : : ; Zp ) then the corresponding pdf equation for pr (t; z1 ; z2 ; : : : ; zp ) is unclosed since it involves an external variable, namely Zp+1 9(g(t; z1 ; : : : ; zp )pr ) 9(g(t; Z1 ; : : : ; Zp ; Zp+1 )|Z r = z r pr ) 9p r 9(g(t; z1 ; z2 )pr ) + ··· + + + 9t 9z1 9zp−1 9zp =0 : (86) To obtain a closed model, the external variable Zp+1 must be expressed as a function of the variables contained in the chosen state vector, and the equations for the modelled system have the form with a model written gm for the time rate of change of Zp d Z1 (87a) = g(t; Z1 ; Z2 ) ; dt d Z2 (87b) = g(t; Z1 ; Z2 ; Z3 ) dt .. . (87c) d Zp = gm (t; Z1 ; : : : ; Zp ) : dt
(87d)
4. Stochastic di-usion processes for modelling purposes The purpose of the present section is to show how stochastic processes can be used in applied situations for modelling issues. Indeed, we have seen in the previous section that the practical need to limit ourselves to reduced descriptions results in unclosed pdf equations. To obtain closed equations, the disregarded degrees of freedom may be replaced by stochastic models. The objective in this section is to try to clarify what is meant when a stochastic process is written to replace a real physical process. This is not always an easy question, though there are some situations when such a move is clear. For example, if we are dealing with a mechanical system subject to an external force F(t) which uctuates rapidly with a variance 2 (t) around a mean term Fd (t), then the obvious model is to write the equivalent of Newton’s law as d Xt (88) = F(t) ⇒ d Xt = Fd (t) d t + (t) d Wt : dt However, the situation is perhaps less clear when we are dealing with internal degrees of freedom. The methodology is thus detailed in the rest of this section starting with a ‘simple’ example.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
35
4.1. The shift from an ODE to a SDE Let us consider the case of a system Xt whose time rate of change is Yt d Xt (89) = Yt : dt We consider that we are dealing with stochastic processes (due, for example to random initial conditions) which are di8erentiable and can thus be handled with normal calculus rules. This gives t dX 2 (t) Y (t)Y (t ) d t : (90) =2 dt 0 If we consider, for the sake of simplicity, Y (t) as a stationary process and introduce its autocorrelation Ry (s) deAned by Ry (s) = Y (t)Y (t + s)= Y 2 , we can write t dX 2 Ry (s) d s : (91) = 2Y 2 dt 0 The important scale in that reasoning is the integral time scale of Y (t), say T , which is deAned as the integral of the autocorrelation ∞ T= Ry (s) d s : (92) 0
This time scale is a measure of the ‘memory’ of the process. If we consider time intervals s small with respect to T , successive values of Y (t) are well correlated. On the other hand, successive values of Y (t) over time intervals that are large with respect to T are nearly uncorrelated. Therefore, in this second limit, we have t for t T; Ry (s) d s ∼ T ⇒ X 2 2Y 2 T × t (93) 0
that is the mean square of X (t) varies linearly with the time interval, here t. This is the ‘di8usive regime’. It should be noted that this regime is always reached (for long enough time spans) and that, once it is reached, the behaviour of X 2 does not depend on the particular form of Ry (s) but simply on two mean quantities, namely the variance and integral time scale of Y (t). This reasoning is certainly not new. Applied to the position and velocity of a uid particle, this point was described by Taylor in 1921 and has been detailed in most textbooks. However, we are not simply interested in reformulating known results concerning the statistics of X (t) but in modelling the instantaneous trajectories. Indeed, if we assume that the trajectories of X (t) are continuous, the previous result suggests that, in the range t T; X (t) can be seen as a Wiener process, that is undergoing a random walk. The previous behaviour is obtained with Anite time di8erence and by Arst introducing T and then making t or Qt large enough. The reasoning can be reversed to reveal what the introduction of a white noise means. We still consider Xt whose time rate of change is Yt . Let us consider that there is a separation of scales: we introduce a time step Qt ∼ d t representing the time interval over which we observe the process Xt . This time increment d t is therefore assumed to be small with respect to a characteristic time of Xt . Nevertheless, we assume that the integral time scale of Yt ; T , is very small with respect to d t. Thus, Yt is a fast and rapidly changing
36
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
variable. Actually, we would like to take directly the limit T → 0, since d t is assumed to be arbitrarily small. Yet, if we take that limit, assuming that Yt is a normal process having a Anite variance, Eq. (93) shows that the e8ect of the uctuations of Yt vanishes completely. Consequently, to retain a Anite limit when T → 0, we are forced to consider that Y 2 becomes arbitrary large in the sense that 2 Y → +∞ (94) such that Y 2 T → D ; T →0 where D is a Anite constant. In that case, the modelling step consists in replacing the di8erentiable process Yt by a white noise and writing that Xt becomes a di8usion process deAned by the SDE √ d Xt = 2D d Wt ; d Xt (95) = Y (t) → T →0 D = lim Y 2 T : dt T →0
By making this step, Xt becomes a Markov process since the memory of Yt becomes inAnitesimally small. It also implies that some ‘information’ has been lost (the information associated to Yt ) in an irreversible way. The signiAcance of this modelling step can be further clariAed by writing the consequences in the pdf equation. If p(t; y) is the pdf associated to the process Xt , we have 9p 92 p 9Yt |Xt = y 9p → =− =D 2 T →0 9t 9t 9y 9y
(96)
which shows that we have in fact introduced a ‘transport coe@cient’, namely D. The discussion above is presented in the framework of continuous-time stochastic processes, and to be put on Arm mathematical grounds the limit expressed in Eq. (94) is required. On a discrete time basis, the time scale T of Yt does not have to go exactly to zero. What is required is that this very time scale be small with respect to the time step which is the reference time scale Qt we have introduced right at the outset. It is important to realize that in practice the introduction of a white-noise term is a relative notion. With regard to one time scale, another process is assumed to vary ‘su@ciently quickly’. Therefore, the details of this fast process are not crucial: the wild variations can be expressed by a Wiener increment. Yet, the eliminated fast process leaves its trace through its variance and integral time scale which deAne the transport coe@cient D. Using a discrete representation of Xt , this step can be expressed by t+Qt t+Qt √ QX (t) = X (t + Qt) − X (t) = Ys d s → QX (t) = 2D d Ws : (97) t
T Qt
t
4.2. Modelling principles All the necessary elements have been given in the above example and can be developed in a more complex context to propose a general methodology. The idea is that introducing a local closure in an (open) set of equations means a Markovian approximation. Such an approximation
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
37
can be justiAed by a coarse-graining procedure, that is by observing the system on ‘large enough’ time intervals. This is precisely what we did in the previous example by taking not-too-small d t in order to disregard the ‘information’ related to Yt and to retain only its e8ects on Xt through the coe@cient D. The success of such a procedure will therefore rest upon a satisfactory choice of the ‘size of the grain’ (in practice a time or length scale) and upon a separation of scales as explained in Section 3. Let us build on these ideas in a complex situation to help us select the proper degrees of freedom to retain in the state vector. As it was explained in the previous sections, for the case of N interacting particles, and even if we limit ourselves to characterizing the statistical behaviour of one particle (one-point pdf), we still have a huge (maybe inAnite) number of degrees of freedom. We could limit ourselves to the position of the particle, say Xt , or include its velocity to have (Xt ; Ut ), or also its acceleration (Xt ; Ut ; At ) and so on. Using the language of Statistical Mechanics or of Synergetics [23,24] the principle is to introduce Arst a reference scale which in our example with one particle would be a reference time scale d t. Then, the degrees of freedom written as Zt = (Z1 ; : : : ; Zn ) are classiAed with respect to that scale as slow and fast variables, (Z1 ; Z2 ; : : : ; Zn ; : : :) ; ↑ reference scale A slow variable is a variable whose integral time scale T is greater than the reference scale d t while fast variables are those with an integral time scale $ smaller than the reference scale, $d t T :
(98)
The guiding principle is then to retain only the slow modes or variables in the state vector used to build the model and to ‘eliminate’ the fast ones. The latter modes are eliminated by expressing them as functions of the slow ones. This is called the slaving-principle [23] and is in fact an equilibrium hypothesis. The fast modes are assumed to relax ‘very rapidly’ to equilibrium values or distributions which are determined or parameterized by the values taken by the slow modes. This corresponds to sorting out the degrees of freedom in terms of solutions of transport equations and local source terms. The slow modes (Z1 ; Z2 ; : : : ; Zd ) that are kept in the state vector will satisfy di8erential equations while the fast ones (Zd+1 ; Zd+2 ; : : :) will be given by algebraic relations. In uid mechanics applications, statistics on the slow modes will be solutions of transport equations while statistics on the fast modes will appear as local source terms. Of course, this procedure will be successful if there exists a clear separation of scales between the integral time scales of the slow modes and of the fast ones. This was indeed the case in the previous example and this allows to replace the fast modes by white-noise or increments of Wiener processes. In the general case, there is no such clear-cut separation and replacing the fast variables by white-noise terms appears as a less justiAed approximation. However, the interest of this principle is at least to provide a convenient and coherent framework and to suggest in practice which variables have the ‘best chances’ to be replaced by a model.
38
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
4.3. Example for typical stochastic models A typical example of this reasoning is the historical case of a Brownian particle. This example was already used in Section 3 to illustrate the pdf hierarchy with respect to increasing one-particle state vectors. We can return to that case and go one step further by introducing speciAc models following the general methodology above. The Arst and simplest description retains only the position of the particle (that was Einstein’s point of view) and with that choice of the state vector Zt =(Xt ), the particle velocity is an external variable and has to be eliminated to obtain a closed model, as already explained in Section 3. When a large enough time step Qt or d t is used, the particle velocity can be regarded as a fast variable and the resulting stochastic model for Brownian particle location is expressed by √ d Xt (99) = Ut → d Xt = 2D d Wt : dt That procedure implicitly assumes that the time scale of the particle velocity Ut , say TU , is small with respect to d t. The corresponding pdf equation is a simple di8usion equation in sample-space (identical to a heat equation) 9p 92 p =D 2 : 9t 9y
(100)
The correlation between successive particle locations is given by Xt Xs = min(t; s) :
(101)
In the Einstein’s picture, particle velocities do not exist. If this Arst picture is believed to be too crude, one can include the velocity of the particle in the state vector that becomes then Zt = (Xt ; Ut ) (Langevin’s point of view). Now, the particle acceleration At becomes an external variable that has to be eliminated. The model proposed by Langevin is written as d Xt d Xt = Ut d t = Ut dt → (102) √ Ut d Ut d U = − d t + K d W t t = At T dt and the corresponding pdf equation for p(t; y; V ) is
1 92 [Kp] 9p 9p 9 1 : (103) +V = Vp + 9t 9y 9V T 2 9V 2 The correlation between successive particle velocities is now given by KT −(t+t )=T KT −|t−t |=T Ut Ut = U02 e−(t+t )=T − e + e : (104) 2 2 When we consider times both long enough with respect to the initial time of the process, the form of the correlation takes the simpliAed expression KT −|t−t |=T t; t T Ut Ut = e : (105) 2
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
39
This reveals that the time scale T used in the stochastic velocity equation is the time scale of particle velocity correlations since ∞ −|s|=T RU (s) = e and thus RU (s) d s = T : (106) 0
The Langevin model has better support if the acceleration can easily be replaced by a model. In the case of a Brownian particle, the acceleration is due to the large number of collisions with uid molecules. Due to the large inertia of the Brownian particle compared to the inertia of uid molecules, we can select a time step which is small with respect to the time scale of particle velocities and yet large with respect to the time scale of uid molecule velocities. The motion of these molecules can thus be seen as a fast and purely random process. The total action of the collisions is written as the sum of two contributions: a purely deterministic one opposed to the Brownian particle motion and a purely random one expressed as a white-noise process. For that precise example the complete form of the Langevin model is written with kB the Boltzmann constant, : the friction coe@cient and ; the uid temperature as d Xt = Vt d t ; d Vt = −:Vt d t +
(107)
2kB ;: d Wt : (108) The Langevin model is really the archetype of stochastic processes for uid dynamical modelling problems and will be extensively referred to in the next chapters. It is therefore important to be aware of its physical justiAcation and, consequently, of its inherent limitations. In the Langevin’s picture, one part of the particle acceleration is taken as a fast process and replaced by a white-noise term. Consequently, information related to the acceleration is lost. If such information is needed, or if acceleration cannot be seen as inAnitely fast, the same procedure can be pursued by shifting the introduction of the necessary model to the time rate of change of At . A useful model can be written as d Xt = Ut d t Ut Ut (109) At = − +
The value of <(t) at time t depends upon the past values. In the acceleration-based model, the process (Xt ; Ut ) is not Markovian anymore. The corresponding pdf equation for p(t; y; V; ) is now
9p 9p 9 1 9 1 9 1 92 [Bp] : (111) +V = Vp − [p] + p + 9t 9y 9V T 9V 9 $ 2 9 2
40
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Far from the chosen origin of time, the correlation between successive uid particle velocities is expressed by B$ T$ 1 $ U (t)U (t + s) = × × × e−s=T − e−s=$ (112) 2 1 + $=T (1 − $=T ) T and the integral time scale of the uid velocities is equal to T + $. Finally, it is also instructing to consider the interplay between these di8erent pdf descriptions. In the last model, the pdf equation for p(t; y; V; ) is closed. From the knowledge of this pdf, information concerning only particle location and velocity can be retrieved by integrating over the extra variable <, p(t; y; V ) = p(t; y; V; ) d ; since p(t; x; V ) is simply the marginal of p(t; x; V; ). From the pdf equation for p(t; y; V; ), Eq. (111), the equation satisAed by the marginal is readily obtained. The integration yields
9p 9p 9 1 9 (113) +V = Vp − [<|(y; V )p] ; 9t 9y 9V T 9V where
<|(y; V )p(t; y; V ) =
=
p<|(x; U ) ()p(t; y; V ) d
(114)
p(t; y; V; ) d
(115)
=<(U − V )(X − y) :
(116)
However, the pdf satisAed by p(t; y; V ) is now unclosed. To obtain a closed form requires to express the mean conditional expectation of < by a function of (y; V ). In general, additional information or assumptions must be input at that stage. 5. The physics of turbulence In this section, we discuss the physics of single-phase ow turbulence. The core of the present work is not directly devoted to this problem, but to the statistical properties of discrete particles. However, since these discrete particles are transported and dispersed by the turbulence of the carrier uid, their statistical properties are strongly in uenced or governed by the underlying turbulent ow. It is then important to detail the physical characteristics of turbulent ows. Turbulence is a very di@cult subject. It is also a subject where di8erent people may actually be following di8erent objectives. Quoting Lesieur [3], this is a subject where people may completely disagree (sometimes in violent terms) on what the problem is. In line with this statement, the point of view adopted in this work is to follow a ‘middle-road approach’ between a more theoretically-inclined point of view and a more engineering-inclined point of view. On the one hand, this section does not claim to present a comprehensive overview of the physics of turbulence and of all theoretical issues. On the other hand, some of these issues are not swept aside on the grounds that they do not concern ‘engineering’ aspects of the problem. The
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
41
presentation remains as general as possible and, as such, this section is the last section devoted to background issues (as well as Sections 2– 4). It provides a link with the following sections which will be devoted to the developments of speciAc models. With respect to these points, the precise aims of this section are (i) to provide a background on the physics of turbulence and in particular on the classical Kolmogorov theories and the concept of the energy cascade, (ii) to describe speciAc results, such as Lagrangian statistics, that will be used in the developments of PDF models in the following sections, (iii) to present and discuss some of the known deAciencies in the classical theories and to introduce recent theoretical issues or results, such as small scale intermittency or coherent structures. These points will help to clarify what will be captured by proposed models, what is explicitly left out and also what could be captured by a certain formalism, (iv) to try to somewhat reconcile theoretical and practical points of view by indicating how some theoretical Andings or concepts could be used in approximate descriptions and then in simple models, (v) to emphasize the physical aspects of turbulence which make the problem inherently di@cult to understand and to model in a satisfactory way. The literature on the subject is immense and impossible to cite in detail. Complementary aspects can be found in several textbooks on turbulence [3,5,25 –27] and in recent reviews which provide up-to-date lists of references [28,1]. 5.1. The turbulence problem The starting point is provided by the Navier–Stokes equations supplemented with conservation equations for a set of scalars. These are Aeld equations (the di8erent Aelds are density (t; x), pressure P(t; x), velocity U(t; x) and scalars (t; x), where x represents the coordinates in physical space) 9 9(Uj ) =0 ; (117a) + 9t 9xj 9Ui 9Ui 1 9P 92 Ui =− + 2 ; (117b) + Uj 9t 9xj 9xi 9 xj 9 l 9 92 (117c) + Uj l = > 2l + Sl () : 9t 9xj 9 xj The set of scalars is a compact notation that includes in a vector the mass fractions m: (: = 1; : : : ; Ns ) of the Ns species that compose the reactive mixture and the enthalpy (or the temperature T ) of the mixture, thus = (m1 ; : : : ; mNs ; T ). This allows us to express the equation of state that gives the mixture density as a function of the scalars for low Mach number ows, where pressure does not in uence density signiAcantly, as = (). The conservation equations for the di8erent scalars involve reactive source terms, Sl for the species l, which depend upon the whole set of species mass fractions and upon the enthalpy or temperature. With the general set of scalars , the dependence of the source terms on the variables can be written as Sl = Sl ().
42
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
The basic equations are known. However, we are dealing here with fully turbulent ows. In that case, it is known that the solutions of the Navier–Stokes equations exhibit uctuations over a wide range of scales. It will be shown later on in Section 5.3, that the number of degrees of freedom is of the order of Re9=4 for a Reynolds number Re that is typically 105 –108 . The existence of such of wide range of scales, and of the acute sensitivity of turbulent ows to small perturbations in initial and boundary conditions (which are never known absolutely) explain the search of a statistical description of turbulent ows. Turbulence is such a vast subject that many di8erent objectives can be pursued. Broadly speaking, the ‘turbulence problem’ is the search for a reduced statistical description. In particular, from a modelling point of view, the ‘turbulence problem’ is to come up with a tractable statistical model. This implies reasonings in which statistical quantities are manipulated or compared. Of Arst importance for statistical descriptions are the characteristic scales which are Arst deAned before going into the details of Kolmogorov theory. 5.2. Characteristic scales Integral time scales have already been introduced in Section 4 through the discussion of simple di8usive behaviour. In this section, we deAne the characteristics time and length scales which are of importance for turbulence modelling and, in the process, we introduce notations that will be used in the rest of the paper. In the statistical approach to turbulence, all variables (velocity, pressure, temperature, etc.) are regarded as random functions or random Aelds. Yet, a fundamental aspects of turbulence with respect to other random phenomena is that the values of each random variable (say velocity) at di8erent locations or at di8erent times are not independent. They are usually correlated or, in other words, turbulence has a non-zero memory. These non-zero memories are best quantiAed by the use of the autocorrelation functions and of the corresponding integral scale. In this section, we limit ourselves, for the sake of simplicity, to homogeneous isotropic stationary turbulence. If we Arst consider a Lagrangian point of view, and record the successive velocities in time of a marked uid element, say Ua (t), where the index a indicates that the particle starts from location a at time t = 0, the autocorrelation coe@cient of the Lagrangian uid particle velocity is then deAned as RL (s) =
Ua (t) Ua (t + s)
u2
:
(118)
In this equation, u2 stands for the constant velocity variance. The autocorrelation coe@cient does not depend on absolute times but on time di8erences since we are considering homogeneous stationary turbulence. A typical form of RL (s) is shown in Fig. 3. As already indicated in Section 4.1, the value of the autocorrelation coe@cient RL (s) for a time lag s is an indication of the degree of correlation between the two velocities Ua (t) and Ua (t + s). For small s, RL (s) 1 and the two velocities are strongly correlated while for long enough time intervals, RL (s) 0 and the two velocities are nearly uncorrelated. A rough measure of the time interval over which velocities remain typically correlated is the integral
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
43
Fig. 3. Typical form of the autocorrelation coe@cient.
time scale deAned as +∞ TL = RL (s) d s :
(119)
0
As shown in Section 4.1, this characteristic timescale is the key parameter for turbulent di8usion. The deAnition of the integral time scale TL allows the notion of ‘long enough time intervals’ used above to be speciAed: a time interval s is long enough for the successive velocities to be uncorrelated when s is large with respect to the integral time scale, sTL . Another scale can be introduced as the curvature at the origin, see Fig. 3. 2 d RL (s) 2 s2 = − 2; RL (s) = 1 − 2 sTL : (120) 2 ds # # s=0 The Taylor scale, #, is the integral over which velocities are strongly correlated (that point will be discussed again in Section 5.3 for spatial velocity di8erences). Since the process Ua (s) is stationary, # represents also a measure of the acceleration variance. Indeed, we have
2 2 2u2 d Ua (s) d Ua (s ) d Ua (s) 2 2 d RL (s − s) 2 d RL (s) = u ⇒ = − u = : ds d s d s d s ds d s2 #2 s=0 (121) The exact form of the autocorrelation coe@cient RL (s) does not play a role for the long-time di8usive behaviour. However, for other concerns (in particular for the expression of x2 (t) when t ∼ TL ), the shape of RL does play a role. Little theoretical information is available but three requirements can be stated |RL (s)| 6 1;
d RL (s) ds
RL (0) = 1;
RL (∞) = 0 ;
(122a)
d 2 RL (s) d s2
s=0
=0 ;
(122b)
s=0
60:
(122c)
44
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 4. DeAnition of the longitudinal and transversal directions.
A simple form that is assumed is an exponential formula, RL (s) = exp(−s=TL ). For example, this expression results from a Langevin stochastic model, see Section 4.3. It is obvious that this expression does not respect the last constraints, Eqs. (122b) and (122c). Nevertheless, these constraints are governed by the behaviour of RL (s) very close to the origin. Away from the origin, the exponential form is acceptable and has been conArmed by both experimental and numerical studies for the case of homogeneous turbulence [29 –31]. The question of the form of the autocorrelation coe@cient and of the limitations of the exponential expression are taken up again and discussed in Sections 6.8 and 7.5.3. Similar reasoning can be applied to the Eulerian velocity Aeld. If we consider the velocities at two di8erent locations and at two di8erent times, we can deAne the Eulerian space–time correlation tensor as RE; ij (r; t) = Ui (x0 ; t0 )Uj (x0 + r; t0 + t) :
(123)
The space and time dependence are generally considered separately. We thus Arst deAne the Eulerian tensor at the same point by RE; ij (t) = Ui (x0 ; t0 )Uj (x0 ; t0 + t) :
(124)
This directly introduces the Eulerian time scale which represents the memory of the turbulent velocities seen by an immobile observer +∞ 2 TE; ij (x0 ) = u RE; ij (t) d t : (125) 0
In an analogous way, the spatial Eulerian tensor is deAned as the correlation between velocities at the same time but at di8erent locations RE; ij (r) = Ui (x0 ; t) Uj (x0 + r; t) :
(126)
This is the quantity usually considered and most analyses are limited to the spatial correlation. For homogeneous isotropic turbulence, the form of the tensor RE; ij (r) can be further developed. Indeed, it can be shown that the only non-zero correlations are the correlations along the longitudinal and transversal directions with respect to the separation vector r, see Fig. 4.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
45
Fig. 5. Shapes of the longitudinal and transversal correlations.
In a general system of reference, the Eulerian tensor has the form f(r) − g(r) 2 RE; ij (r) = u g(r) ij + ri rj ; r = |r| ; |r|2
(127)
where f(r) is the longitudinal autocorrelation coe@cient and g(r) the transversal coe@cient. This relation results from symmetry conditions. Additional information is provided by dynamical constraints. For example, the continuity equation implies that 9RE; ij (r) =0 : (128) 9 ri Application of Eq. (128) in isotropic turbulence yields the di8erential equation r d f(r) g(r) = f(r) + : (129) 2 dr As for the Lagrangian velocity autocorrelation function, few requirements can be stated. The conditions listed above for RL (s) are still valid and it can be shown that the transversal correlation function g(r) must contain negative loops, since mass conservation across a plane implies that +∞ g(r) d r = 0 : (130) −∞
On the other hand, experimental and numerical studies tend to suggest that the longitudinal autocorrelation function f(r) does not present such negative loops. A sketch of both correlations shapes is given in Fig. 5. Numerical simulations support an exponential form for f(r) except in the vicinity of the origin (as for RL (s)). Hence, f(r) and g(r) can be approximated by the relations f(r) exp(−r=LE; ) ; r g(r) 1 − exp(−r=LE; ) : 2LE; In these equations, LE; stands for the Eulerian longitudinal length scale deAned by +∞ LE; = f(r) d r : 0
(131a) (131b)
(132)
46
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 6. Sketch of the Kolmogorov cascade.
Similarly, a transversal length scale is introduced by +∞ LE; ⊥ = g(r) d r : 0
(133)
An important result is that, even in isotropic turbulence, the Eulerian tensor is not isotropic and that correlations depend upon whether the longitudinal or transversal direction is considered. The two length scales are related by LE; ⊥ = LE; =2. 5.3. Kolmogorov theory Ever since their Arst formulation by Kolmogorov, the set of hypothesis that form what is now referred to as the ‘Kolmogorov theory’ has been the cornerstone of turbulence modelling. This picture or description of turbulence has indeed acquired a central role, acting as a reference theory in most analyses whether the point is to conArm its predictions or to pinpoint its limitations. We limit ourselves to presenting only the salient points of the Kolmogorov theory, and we emphasise mostly the applications that are of particular interest for the development of Lagrangian stochastic models to be considered in Sections 6 and 7. A comprehensive presentation of the Kolmogorov theory can be found in Monin and Yaglom [25], and the theory is discussed in several textbooks on turbulence [5,3,27]. The Kolmogorov theory is a phenomenological description of turbulence based on the idea of a cascade of energy from large to small scales through a range of what is called inertial scales where energy is merely transferred to smaller scales without creation nor dissipation. The Kolmogorov picture is one of quasi-equilibrium where + is assumed to represent at the same time the rate of production of kinetic energy at the large scales, the rate of transfer in the inertial range and Anally the rate of viscous dissipation at the end of the turbulence spectrum though they are di8erent phenomena. It is represented in Fig. 6 where three di8erent ranges have been separated: the production range (P.R.) where kinetic energy is produced by large scale motions, the inertial range (I.R.) where kinetic energy is merely transferred to smaller scales and the dissipative range (D.R.) where kinetic energy is dissipated into heat by viscous forces.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
47
If we call u a typical uctuating velocity, the characteristic large scales are expressed by L∼
u3 ; +
T∼
u2 : +
(134)
The Arst hypothesis in the Kolmogorov theory is to say that, at high Reynolds numbers, turbulence can be seen as locally isotropic at least for the range of scales ‘su@ciently’ far from the large energy-producing ones. The local statistics of quantities belonging to that range, for example velocity increments in space and time v($; r) = U(t + $; x + r) − U(t; x) relative to the constant motion of one uid particle at the velocity U(t; x) (to have a Galilean transformation), are actually independent of the reference velocity U(t; x) and are uniquely deAned by the two parameters + and and the corresponding time and space increments ($ and=or r) which deAne the local quantity itself. The characteristic scales at which viscous damping gets the upper hand over turbulent uctuations are thus given from simple dimensional analysis by 3 1=4 1=2
@= ; $@ = : (135) + + The ratios of the smallest to the largest scales can be expressed from the Kolmogorov scales and from the relation for the integral length scale L given above in Eq. (134) @ ∼ Re−3=4 ; (136a) L u@ ∼ Re−1=4 ; (136b) u $@ ∼ Re−1=2 : (136c) T With the expression of the inner scales, the inertial range can now be deAned precisely as the scales (r; $) such that with r = |r| @r L;
$@ $T :
(137)
The second hypothesis states that in that inertial range molecular viscosity becomes irrelevant and that statistics of U($; r) = v($; r) depend only on $; r and +. Once the large and small scales have been introduced, we can give another description of the energy cascade in a more quantitative way. If we already adopt an Eulerian point of view (that will be pursued just below) and deAne eddies as ‘coherent’ turbulent motions localised in a region of size r with characteristic velocities u(r) and time scales $(r), we can now express these scales as functions of the large and small scales. The key notion is the idea of a constant transfer of energy, + which implies that relations similar to Eq. (134) are valid not just for the largest scales but at any scale r. This yields the following relations: u(r) = (+r)1=3 = u@ (r=@)1=3 ∼ u(r=L)1=3 ;
(138a)
$(r) = r=u(r) = (r 2 = +)1=3 = $@ (r=@)2=3 ∼ T (r=L)2=3 :
(138b)
48
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
The Kolmogorov picture implies that, in the inertial range, the velocity scales u(r) and the time scales $(r) decreases as we consider smaller and smaller scales r. Then, for a small enough scale r, the time scales $(r) are small enough compare to the time scale of the large scale motion $(L) ∼ L=u, so that these eddies can be regarded as fast processes or variables in the spirit of the discussion in Section 4. Following the general discussion given in Section 4.2, these small eddies will relax very rapidly to their equilibrium values while the large scale, which are the energy-containing scales, have not changed very much and therefore while the ux of energy remains approximately constant. This approximatively constant energy ux acts therefore as an ‘external constraint’ for the small scales. We then obtain the physical notion of the universal equilibrium range which is an important aspect of the Kolmogorov description. Furthermore, in the cascade process and the intrinsically chaotic size-by-size production of smaller eddies, directional information is lost and the resulting small scales are assumed to be statistically isotropic. However, the consequence of the Kolmogorov theory is not as nice as it would seem for the purpose of turbulence modelling and for most practical purposes. Indeed, the universal behaviour which provides the main justiAcation for the search of a (universal) model, is only valid for the small scales. Yet, most of the turbulent energy is contained in the large scales or within a range of scales which are comparable with the integral length scale re ∼ L. These scales have a size comparable to the geometry of the ow and are still very much a8ected by the anisotropy of the ow. Furthermore, their characteristic time scale $(re ), which is thus of the order of magnitude of the integral time scale T , is not small compared to the time scale of the mean ow L= U . For example, in typical shear ows such as mixing layers or shear ows, this characteristic time scale is actually four times higher than the mean- ow timescale. These scales contribute mostly to transport phenomena but do not have a universal form produced by statistical equilibrium. This simple conclusion illustrates the inherent di@culty of practical turbulence models which often have to express the e8ects of these scales through constitutive laws. This Arst form of the Kolmogorov theory, published in 1941, is currently referred to as K41 in the literature and this notation will be kept here. Before considering more in detail speciAc applications, it is worth emphasising that Kolmogorov theory is essentially a Lagrangian concept since predictions are made for Aelds relative to the motion of one uid particle, and that predictions concern the form of N -point distributions in the four-dimensional space (x; t) and include both space and time. Most presentations of Kolmogorov theory limit themselves to space N -point distributions and leave out temporal issues. Yet, the predictions for correlations in time of various quantities are of great importance for the Lagrangian models to be considered in Sections 6 and 7. 5.3.1. Eulerian statistics of velocity diCerences Kolmogorov hypotheses are mostly applied to space correlations, that is for the statistics of Ur =v(0; r)=U(t; x +r) − U(t; x). Statistics often considered are the velocity–structure functions deAned as F n (r) = (Ur )n ;
(139)
which can have di8erent forms depending on the orientation of the velocity component with respect to the separation vector r (the longitudinal and transversal components), see Fig. 4.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
49
When the separation distance lies within the inertial range, we can apply Kolmogorov hypotheses. First of all, since turbulence is locally isotropic, the results given in the previous subsection still hold. The Arst-order moment is zero and the second-order tensor is expressed with the help of the longitudinal and transversal correlations as in Eq. (127). We thus consider separately the two Eulerian velocity structure functions along the longitudinal and transversal directions, D and D⊥ , respectively. Kolmogorov theory predicts that when @r L D = (U; r )2 = C(+r)2=3 ;
(140a)
D⊥ = (U⊥; r )2 = 43 C(+r)2=3 ;
(140b)
F3 = (U; r )3 = 45 +r :
(140c)
The Kolmogorov prediction for the second-order moment is often re-expressed in terms of the energy spectrum (Ur )2 ∼ (+ r)2=3
(Ur )2 ∼ kE(k) ⇒ E(k) ∼ +2=3 k −5=3 ;
(141)
leading to the famous expression of the energy spectrum (k is the wave number). Eqs. (140) already indicate that relative velocities are not Gaussian since odd moments are not zero. By the same reasoning, Kolmogorov hypotheses imply that the n-order velocity structure function has the form Fn (r) = (U; r )n = Cn (+ r)n=3 :
(142)
The immediate consequence of Kolmogorov theory is that the non-dimensional structure factors are independent of the separation distance r and are universal constants Fn (r) (U; r )n fn (r) = 2 = = :n : (143) (F (r))n=2 (U; r )2 n=2 An important consequence of the above scaling is that the probability distributions of the normalized (with unit variance) variable U; r =(+ r)1=3 should not depend on the scale r when r lies in the inertial range and should be universal. Finally, the form of the second-order moments in the inertial range can be used to bring out another signiAcation of the Taylor length scale, #. From the deAnition of the rate of dissipation of the turbulent kinetic energy, 9U 2 +
(144) 9r and the relation between the variance of the velocity derivative and the Taylor scale, see Eq. (121), we have + ∼ u2 =#2 and thus that # ∼ L1=3 @2=3 . From this, we can show that # can be regarded as the averaged length scale of strong correlations when velocities are rescaled to u. Indeed, we have that Ur2 ∼ (+ r)2=3 and thus the dissipation of turbulent kinetic energy at that scale by viscous forces is +r =
Ur2 +2=3 r −4=3 : r2
(145)
50
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
When the turbulent kinetic energy and its dissipation at r are re-scaled we have @ 4=3 Ur2 r 2=3 +r ; : u2 L + r
(146)
The scale at which the relative quantities are of the same order is thus r ∼ L1=3 @2=3 that is when r = #. 5.3.2. Lagrangian statistics of ;uid particle velocity increments The Kolmogorov theory can also be used to predict the more general statistics of the Aeld v (which is in spirit a relative Lagrangian velocity Aeld), and for Lagrangian statistics of a uid particle. Direct application reveals that, if one considers time intervals d t in the inertial range $@ d t T , the successive velocities U (t) and U (t + d t) are strongly correlated. Indeed, we have d UL = U (t + d t) − U (t) = v(d t; 0) ⇒ (d UL )2 ∼ C0 + d t ;
(147)
where C0 is a constant. This implies for the velocity autocorrelation function DL (d t) that U (t)U (t + d t) DL (d t) C0 d t DL (d t) = ∼1− ∼1− ∼1: (148) u2 2u2 2 T For the statistics of uid particle acceleration A(t), the same hypotheses yield that + + A(t + d t)A(t) ∼ (149) ; A2 ∼ dt $@ which implies for the acceleration autocorrelation function RA (d t) that $@ A(t + d t)A(t) RA (d t) = ∼ 1 : 2 A dt
(150)
Consequently, for a given reference time scale d t which belongs to the inertial range, the successive uid particle accelerations are nearly uncorrelated while their velocities are still well correlated. In the spirit of the discussion of Section 3, we can conclude that the uid particle acceleration can be regarded as a fast variable. The Kolmogorov theory suggests therefore that the uid particle acceleration is close to a white-noise process (its spectrum is constant in the inertial range EA (!) ∼ + where ! is the frequency). This is a crucial result for the development of stochastic models either in the single-phase ow case, for uid particles in Section 6.6, or in the two-phase ow case in Section 7.4.2. 5.3.3. Statistics of temporal velocity increments The Kolmogorov theory can actually be applied to numerous variables either of a Lagrangian or an Eulerian form. This includes the statistics of temporal velocity increments which are important, the statistics of pressure correlations, vorticity, etc. They are discussed at length in the reference textbook of Monin and Yaglom [25]. Since the results of Kolmogorov predictions for temporal velocity increments will be used for two-phase ow modelling in Section 7.4.2, they are brie y presented here. For temporal velocity increments, one is interested in obtaining the statistics of the Eulerian velocity di8erence at a Axed point x0 at two instants t0 and t0 + $. For small time intervals $, this velocity di8erence
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
51
can be expressed in terms of the velocity Aeld v deAned in the reference system moving with velocity U (t0 ; x0 ) as U$ v(−U (t0 ; x0 )$; $) :
(151)
However, the situation is in fact more complicated than for the Eulerian increments or Lagrangian di8erences considered before. Indeed, the variable U$ depends explicitly on the reference velocity U0 = U (t0 ; x0 ). Therefore, applying Kolmogorov theory, one can only conclude that there is a conditional probability distribution for U$ . For a given value of U0 , the tensor ∗ which correspond to the longitudinal D∗ = Ui; $ Uj; $ depends on the two functions D∗ and D⊥ and transversal directions, as in Eq. (127). These two functions depend on $ and on r = U0 $ and in the inertial range for ($; r) we have |U0 |2 |U0 |2 ∗ ∗ D = + $ : ; D⊥ = + $ :⊥ ; (152) +$ +$ where : and :⊥ are universal functions which have to be speciAed. The dependence on U0 and the conditional nature of the previous results can be removed by resorting to the Taylor, or frozen turbulence, hypothesis. In this hypothesis, it is assumed that the turbulent uctuations are much smaller than the mean velocity, or U0 U0 . Then, the turbulent uctuations at a Axed point and over the time interval $ can be regarded as the transport of the turbulent uctuation with a constant velocity U0 without distortion. This frozen-turbulence assumption not only removes the dependence on U0 , but allows to write the statistics of U$ in terms of the velocity structure-functions derived for the Eulerian space increments in Section 5.3.1 by replacing r by U0 $. This leads to D∗ = D (U0 $) ;
∗ D⊥ = D⊥ (U0 $)
(153)
and thus in the inertial range to D∗ = C(+U0 $)2=3 ;
∗ D⊥ = 43 C(+U0 $)2=3 :
(154)
5.4. DiIculties and re?nements 5.4.1. High-order statistics Quickly after the publication and di8usion (through Batchelor’s work) of Kolmogorov hypotheses, measurements were performed on particular velocity statistics. The statistics that were measured, and that have been repeatedly analysed are the atness and skewness factors of velocity derivatives. The skewness and atness factors are the third and fourth moments of the velocity spatial derivatives. For the case of the gradient of the longitudinal velocity component along the longitudinal direction, say 9U1 = 9x1 , these factors are deAned by S=
(9U1 = 9x1 )3 ; (9U1 = 9x1 )2 3=2
K=
(9U1 = 9x1 )4 : (9U1 = 9x1 )2 2
(155)
52
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
They are just special cases of high-order velocity-derivative moments which are expressed by (9U1 = 9x1 )n : (156) Mn = (9U1 = 9x1 )2 n=2 According to Kolmogorov theory, for each value of n; Mn is a universal constant. In particular, Kolmogorov theory predicts therefore a constant value of S and K regardless of the Reynolds number above a possible threshold since Kolmogorov theory is applicable for high-Reynolds ows. The early measurements were performed by Townsend in 1947 and are reported in Batchelor [32]. Since then, additional data have been gathered by numerous researchers both experimentally and numerically. Experimental data collected in the atmosphere are already discussed in Monin and Yaglom [25] and further results have been obtained in laboratory ows. Data were compiled in Van Atta and Antonia [33] and are discussed more in detail in the recent review of Sreenivasan and Antonia [28]. With the recent development of DNS, similar results have also been found in numerical simulation [34]. All the results indicate that, contrary to Kolmogorov predictions, the skewness (actually −S which is a positive number) and atness factors increase with the Reynolds number apparently without reaching a limit value. Measurements show that the atness increases from K ∼ 4 at low Reynolds number (for Re# ∼ 40 in grid turbulence) to K ∼ 40 at high Reynolds number (for Re# ∼ 104 in the atmospheric boundary layer) and still seems to increase with higher Re# . A similar trend is observed for −S although the rate of increase as a function of the Reynolds number is slower. Similar discrepancies arise, not just for high-order statistics governed by scales in the dissipative range, but also for statistics pertaining to the inertial range. We have seen that in the inertial range, Kolmogorov theory predicts that the n-order velocity structure functions are given by Eq. (142), and therefore that F n (r) scales as r 2=3 when r belongs to the inertial range. High-order statistics have been measured in detail by Anselmet et al. [35] and have been since then conArmed by numerous works (see the discussion in the review of Sreenivasan and Antonia [36]). Measurements conArm the power-law scaling in the inertial range, that is F n (r) ∼ r −Cn , but the measured exponents Cn di8er from the Kolmogorov values showing a slower increase of Cn with n than predicted by Kolmogorov theory with a marked di8erence from n ¿ 5. The scaling exponents Cn are called anomalous exponents. 5.4.2. Intermittency The discrepancies between Kolmogorov predictions and experimental Andings for high-order statistics of velocity increments are mainly attributed to the phenomena of internal intermittency. Indeed, Kolmogorov theory was rapidly questioned by a remark which is believed to have been Arst formulated by Landau (see [5]). In the K41 picture, the energy transfer (taken as identical to the dissipation rate in the equilibrium range, see Fig. 6) was assumed to be given by the mean dissipation rate +. However, the dissipation rate is directly expressed in terms of the velocity Aeld as
9ui 9uj 2 += + : (157) 2 9xj 9 ui Therefore, + is also a random variable which uctuates around its mean value +. These uctuations may depend upon the Reynolds number and involve large scales. As a consequence,
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
53
statistics in the inertial range may depend on large scales and thus are not universal. For example, if we consider that + takes di8erent values in di8erent regions, we can apply the K41 statistical results based on the value of + that exists in each region. Since the structure functions are not linear functions of + (with the exception of the third-order moment which is thus a quantity on which intermittency has no e8ect), it is obvious that the averaged (on the di8erent regions) result di8ers from the global one built on + since F n (r) = Cn +n=3 r n=3 = Cn +n=3 r n=3
when n = 3 :
(158)
One can note that the distribution of the dissipation rate + does not contradict the existence of power laws in the inertial range but rather the universal nature of the constants that appear in Eqs. (142) and (143). In 1962, Kolmogorov (as well as Obukhov) proposed the so-called re?ned similarity hypotheses, referred to as K62. They basically consist in stating that the statistical laws of K41 remain valid but should be regarded as conditional results for a Axed value of + in a small time–space domain V(t; x) centred around the point x where statistical laws are expressed, that is for +r deAned as 3 +r (t; x) = +(t; x + r) d r : (159) 4(r 3 V(r) Averaged results are then derived by integrating the conditional results for all possible values of the dissipation rate +r . For example, the reAned hypotheses predict that the Eulerian velocity structure functions are given by (Ur )n |+r = + = Cn (+r)n=3
(160)
with Cn a universal constant. The unconditional structure functions are then obtained as F n (r) = (Ur )n = T(Ur )n |+r = +U = Cn +n=3 r n=3 :
(161)
The reAned hypotheses introduce the probability distribution of +r (or equivalently of spatial derivatives) in the picture. Furthermore, from Eq. (157), it is clear that + is a small scale variable and that + and +r are mainly determined by small-scale behaviour. Thus, even though similar expressions are obtained in the K62 theory, the overall picture di8ers slightly from the one-way cascade (from the large to the small scales) in K41: small scale statistics are determined from the +-distribution which in turn results from small scale behaviour. To work out global results, this probability must be input. The earlier measurements by Townsend and reported in Batchelor [32] (see the discussion in Section 5.4.1) indicated that small scales deviate from Gaussianity and, as already presented, all subsequent studies have conArmed this trend. The typical pdf that has emerged for small scale quantities has a typical feature: a higher central core and tails which decrease far less rapidly than for the Gaussian distribution (this point is taken up in the next Section 5.5). The deviation from Gaussianity is expressed in terms of an intermittent factor (through the measured atness factor) which can be seen as the fraction of zero and non-zero values. This corresponds to an uneven spatial distribution of small scale quantities: turbulence activity tends to concentrate in conAned parts of the ow (meaning locally a high dissipation and the existence of smaller scales) surrounded by more quiescent regions. This intermittency becomes more and more marked as the Reynolds number is increased and is therefore related to the non-Gaussian character of small scales and the scale dependence of the probability distribution of Ur .
54
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
In order to take intermittency into account, Kolmorogov proposed in 1962 a third hypothesis. It is assumed that +r = + is a uctuating quantity which is a function of r=L, and by making the same assumption for r = @ that += + is function of @=L. More precisely, Kolmogorov assumed that +r (and thus +) is log-normally distributed with a variance given by L 2 : (162) ln(+r ) = A + D ln r The log-normal distribution is the natural distribution for the result of a cascade process with multiplicative noise. In the reAned picture K62, the probability distributions of the normalized velocity increment U; r =(+r)1=3 ) are no longer universal, but we still expect power-law scaling for the structure functions. Indeed, by using the log-normal assumption of +r the Eulerian velocity structure functions F n (r) scale as F n (r) ∼ r −Cn with scaling exponents given by
1 1 (163) Cn = n 1 − D(n − 3) : 3 6 For the second-order structure function, say for the longitudinal function D (r), the correction is very small since the new exponent is now C2 = 2=3 + 1=36 and thus di8ers from K41 prediction only by 1=36. Similarly, the modiAcation to the −5=3 law of the energy spectrum E(k) in the inertial range is minor. When compared to measured scaling exponents, the reAned predictions perform better than the original K41 ones [35] and acceptable agreement is obtained for values of n up to 10. For higher-order exponents, K62 values depart from measured ones, indicating that the reAned turbulence theory can only claim to be an approximate description (but still a very reasonable one) and is not the ultimate theoretical word on turbulence. In spite of known theoretical deAciencies [5], the reAned hypotheses have rather well come out of recent tests, both experimentally [28] and numerically [37]. So far, the discussion on intermittence has mainly been focused on Eulerian statistics such as the (Eulerian) velocity structure functions F n (r). However, intermittence a8ects also Lagrangian statistics, and these e8ects are particularly relevant to the present work since the Lagrangian point of view will be central in the development of pdf models. In the Lagrangian formalism, the pressure Aeld, or more precisely its gradient, plays an important role. Indeed, by writing the Navier–Stokes equation for a uid particle ai (t) =
d Ui (t) 1 9P + QUi ; =− dt 9xi
(164)
it is seen that the pressure Aeld is closely related to the uid particle acceleration statistics. The pressure is given by a Poisson equation QP = −
9Ui 9Uj ; 9xj 9xi
(165)
which shows that pressure is a8ected by large and small scales. Lagrangian statistics are much more di@cult to measure experimentally and less data is available. Results have nevertheless been obtained [25,38]. Valuable information and insight have come from recent direct numerical simulations [29,39]. These results have also brought out discrepancies with the K41 description of turbulence statistics. According to Kolmogorov K41 predictions, the acceleration variance
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
55
varies as a2 = a0 +3=2 = 1=2
(166)
with a0 being a constant. Yet, in the DNS results of Yeung and Pope [29], it is reported that acceleration variance increases with the Reynolds number Re# as a2 ∼ Re#1=2 : (167) +3=2 = 1=2 This is due to intermittent e8ects, but Kolmogorov reAned hypothesis leads to a variation of a0 as a0 ∼ Re#D3=2 which is weaker than the one observed. The Re#1=2 scaling has been conArmed in another analysis of DNS results [39] where it is also found, for Re#1=2 up to 172, that (∇P)2 scales as Re#1=2 . These issues, of anomalous scaling, small scale intermittency, and thus of intermittency models, have received considerable attention and remain a major issue in most theoretical work. Detailed information and guidelines into the related vast literature on this speciAc subject can be found for example in recent reviews [1,28,40] and in textbooks [5]. Recent developments include for instance multi-fractal formalism [5], ESS (extended self similarity) [41] or log-Poisson statistics [42] among other models. 5.5. Experimental and numerical results Analysis and tests of Kolmogorov predictions were not limited to determining the scaling exponents. These exponents are of key theoretical importance but of less interest in the practice of modelling purposes. The di8erent investigations have also helped to bring out useful information on the pdf of various quantities, both of large and of small scales. These results have been obtained from experiments [25,43– 45] and from numerical simulations [46,47,29,39] which have been particularly useful for small scale quantities and for Lagrangian variables. 5.5.1. Experimental and numerical results on distributions We present Arst one-point velocity pdf, then velocity derivative and velocity increment distributions, and results for pressure-gradient pdfs. Velocity distribution. First of all, both experimental and numerical approaches have well conArmed that, for homogeneous turbulence, one-point velocities have a Gaussian distribution [46]. The normal distribution is the natural reference for systems with large numbers of degrees of freedom as a consequence of the Central Limit Theorem. Batchelor showed in 1953 [32] that the Gaussian nature of some band-limited eddies is equivalent to the independence of their Fourier coe@cients over the corresponding wave number band. The distribution of one-point velocity involves the whole spectrum. Yet, we have seen that the energy contained in each band (or the energy density E(k)) and thus that the pdf re ects mostly the large scale behaviour. These scales are excited by uctuations of boundary conditions which are not controlled and it is therefore not too surprising to And an error law distribution. Another way to describe this result is to say that the large scales do not exhibit (internal) intermittency. This is not the case of free shear ows, such as jets or mixing layers, which are known to be intermittent even on the large scale. However, this intermittency results from the mixing of a turbulent core with
56
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 7. Typical forms of the distributions of velocity increments X = Ur for decreasing values of the separation distance r: of the order of L (D), in the inertial range (C) and (B) and in the dissipative range (A). The curve (A) is identical to the distribution for velocity gradients.
laminar surroundings and is thus due to external forcing or conditions (this intermittency is referred to as ‘external intermittency’). For homogeneous turbulent ows, external intermittency is mainly absent, though this does not mean that large scale structures cannot be found (see below). The Gaussian nature of the large scale part of one-point velocity in homogenous ows was well demonstrated by She et al. [47]. The authors used a Fourier band Altering method to isolate and quantify deviations from Gaussianity of separate ranges (large, inertial and dissipative ranges). They found that not only the full velocity Aeld, but also the ‘inertial’ velocity Aeld follow a Gaussian distribution [47] whereas the velocity Aeld corresponding to the dissipative range displays deviations from Gaussianity with near exponential tails. Velocity derivative distribution. The situation is quite di8erent for velocity derivatives since we are now dealing with the other end of the spectrum (velocity gradients are mainly governed by small scale quantities). All available results, from the early experiments [43] to recent simulations [34], indicate that distributions of velocity gradients or derivatives deviate strongly from Gaussian pdfs. As already indicated in Section 5.4.2 the observed distribution is closer to an exponential law for the tails of the distribution and has a central core with higher probability compared to Gaussian values. When the separate contributions of di8erent energy ranges (large scale, inertial or dissipative) are studied [47], the resulting distributions conArm that velocity gradients are indeed small scale phenomena since the distributions for the full Aeld and for the dissipative range show similar behaviour whereas the distribution obtained for the velocity Aeld corresponding to the inertial range is nearly Gaussian. Velocity gradients are actually special cases of velocity increments Ur (when r → 0), so it is useful to consider the distributions for Ur when r varies from the large scale down to Kolmogorov scales. The outcome is that the distributions vary continuously from a Gaussian law when r is of the order of the large scales to the near-exponential pdf characteristic of velocity gradients when r → @. This is illustrated in Fig. 7 which has been observed in many experiments and simulations [35,44,46,34].
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
57
Similar behaviour has been observed for the velocity increments in time of a uid particle QU$ (t) (a Lagrangian quantity) by Yeung and Pope [29]. For time lags which are large enough, that is large with respect to Kolmogorov timescale ($$@ ), the distribution is Gaussian. When smaller time lags are considered, the distributions change continuously showing an increasingly non-Gaussian behaviour and when $ $@ the distribution presents the typical form known for 9U= 9t ever since the early measurements of Van Atta and Chen [43]. Pressure and small scale invariants. Analysis of the distributions of the one-point pressure has shown a non-symmetric pdf [48,39]. For positive uctuations, the distribution appears to be Gaussian and rather insensitive to variations of the Reynolds number Re# . For negative uctuations, the form of the pdf tends towards an exponential distribution (or perhaps stretched exponential) and the tails of the distribution extends to larger negative values as the Reynolds number increases. The pdfs for pressure-gradients have been computed by Gotoh and Rogallo [39]. The resulting pdfs are isotropic and have the same form as velocity gradients or small scale quantities. The distributions di8er markedly from the Gaussian one and have exponential tails (or stretched exponential forms). Therefore, the pressure-gradient Aeld appears to be very intermittent with a variance that increases with Re# as already indicated in Section 5.4.2. Three small scale invariants can be formed from the tensor of the velocity gradients Gij = 9Ui = 9xj . By introducing the symmetrical Sij and anti-symmetrical tensors Rij 1 9Ui 9Uj + ; (168) Sij = 2 9xj 9xi 1 9Ui 9Uj Rij = − ; (169) 2 9xj 9xi we can deAne the three invariants s2 = 2 Sij2 = 2 Tr(SS ⊥ ) = 2 Tr(S 2 ) ;
(170)
!2 = 2 R2ij = 2 Tr(RR⊥ ) = −2 Tr(R2 ) ;
(171)
g2 = Gij2 = Tr(GG ⊥ ) :
(172)
These invariants are directly related to dissipative quantities since + = s2
mechanical dissipation ;
C = !2 enstrophy ; ’ = g2
pseudo-dissipation :
(173)
In homogeneous turbulence, the mean values of these dissipative quantities are identical + = C = ’. Then, by manipulating matrix identities, it can be shown that +; ’ and C satisfy the relations 1 ’ = [+ + C] ; 2
(174)
58
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
+−’=
9Ui 9xj
9Uj 9xi
:
(175)
Kolmogorov reAned hypotheses (K62) assumes that small scale quantities follow log-normal distributions. That hypothesis is often tested by calculating the pdf of the total or mechanical dissipation +, for which the agreement with log-normality is moderate [46,37]. However, in an interesting analysis, Yeung and Pope [29] showed that the distributions of the three small scale invariants +; ’ and C were not identical. It was found that while the pdfs of + and C di8er from the log-normal pdf, the pdf of the pseudo-dissipation ’ on the other hand is very close to being log-normal [29]. This could be related to the di8erent physical meanings of the small scale invariants (see below and Section 5.6.2). Along the same line, a very interesting analysis of conditional pdf of velocity increments was proposed by Gagne et al. [45]. Using experimental data, the authors showed that when conditional velocity increments U (r)|+(r) = +0 are considered, then the observed distributions are Gaussian regardless of whether the separation distance r lies in the inertial range or in the dissipative range. This result was obtained only when +(r) is deAned as the energy transfer rate and not as the energy dissipation rate. From the results of Gagne et al., it appears that intermittent non-Gaussian statistics can be attributed essentially to uctuations of the energy transfer rate. A strong point is that universal Gaussian behaviour can be obtained from quantities that essentially represent energy transfer rate and not energy dissipation rate (though the two have identical mean values). 5.5.2. Spatial structures The presentation and the discussions developed up to now in this section have followed the statistical approach to turbulence. Since the early developments and Kolmogorov theory in particular, this statistical description has been the point of view mainly adopted. Yet, other points of view are possible, among which an interesting one is the geometrical approach. In this approach, one attempts a geometrical description of turbulent ows as a collection of spatial structures. The underlying aim is to try to isolate elementary structures that would play for turbulent ows the role of molecules in classical statistical physics. This point of view was really initiated when it was realized that large structures can exist in turbulent ows, in particular in free shear ows such as mixing layers [49]. These structures are, however, large scale structures strongly related to the geometry of the ow and to external forcing. It is then di@cult to come up with a universal theory for them. Another impetus for the geometrical approach was given by the discovery of the existence of typical small scale structures which are the result of Navier–Stokes dynamics and may present universal characteristics. These small scale structures have been mostly revealed by direct numerical simulations which have proved invaluable tools for their analysis [50,34,51]. Recent experiments have supported their existence and have helped to assess their characteristic features [52,53]. It was found that intense events related to small scales, especially small scale vorticity, were not evenly and randomly distributed in the ow domain, but organized in the form of vortex Alaments. These Alaments are also referred to as vortex tubes or worms. The analysis of these objects is still going on since open questions remain (see the discussion in the introduction of the article by Jimenez and Wray [51]). We simply summarize here present knowledge, drawing
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
59
in particular from the work of Jimenez and coworkers [34,51]. These authors have tried to determine the statistical properties of these geometrical objects. From the recent analysis of DNS results [51], it appears that • a turbulent ow contains small scale intense structures which have the form of long tubes, • these tubes have a radius of the order of the Kolmogorov length scale @ and a length of the
order of the integral length scale L,
• the internal velocity di8erences within the tubes scale as u the large scale characteristic
velocity uctuations, • the tubes occupy a volume fraction which seems to scale as Re#−2 (or at least as Re#a with a ∼ 2). Other typical properties can be deduced from the ones above. The typical vorticity of the tubes scale as u u # # !∼ = (176) = ! ∼ !Re#1=2 @ # @ @ with ! being the rms magnitude for the ow as a whole. With respect to the existence of these tubes, the key questions are: how relevant are they for the dynamics of the ow? What role do they play in the energy budget? How are they related to intermittency? For these purposes, the crucial step is the evaluation of the volume fraction :vt occupied by the vortex Alaments. From the estimation that :vt ∼ Re#−2 , it appears that the vortex tubes give negligible contributions to the kinetic energy as well as to the total dissipation and the total enstrophy of the ow. In terms of statistics, the e8ects of the vortex Alaments on velocity gradient statistics and on velocity structure functions are negligible for low-order statistics (say for n 6 4) but dominant for high-order statistics (n ¿ 4), as discussed in the conclusion of Jimenez and Wray [51]. The interesting conclusion of these works is that the observed deviations of high-order statistics from Kolmogorov predictions could be attributed to the existence of special small scale intense structures, such as vortex Alaments. 5.6. Simpli?ed images of turbulence and Lagrangian models In this subsection, we put forward a simple description of turbulence and we present a Lagrangian point of view in which turbulence is addressed as a N -particle problem. The description tries to include some of the recent ideas or Andings on turbulent ows while remaining simple enough so that tractable models can be derived (or justiAed) from it. This somewhat crude image is therefore meant as a compromise between the complexity of some advanced theoretical concepts in turbulence and the necessary simpliAcations inherent to most models. The purpose of this subsection is to show how some of these concepts can be used even for modelling issues and also to introduce a Lagrangian modelling point of view that will be adopted in the following sections. As such, this subsection provides a link between the theoretical-oriented presentation of the physics of turbulence and the modelling-oriented issues that will be addressed in the rest of this paper.
60
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
5.6.1. A two-fold picture of turbulence We have seen that most of the di@culties in the theory of turbulence (in particular in Kolmogorov theories) come from the phenomena of small scale intermittency. Intermittency is responsible for the deviations of statistics from Gaussianity and can be seen as a manifestation of an underlying order in turbulent ows. The recent developments described in Section 5.5.2 have revealed that this order could be related to intense vortex Alaments. Of course, these vortex tubes are just one of several typical structures that are bound to exist in turbulent ows. They correspond to the most intense small scale vorticity events and it is quite probable that there exists a whole range of structures with di8erent properties. Yet, they suggest the following model picture of turbulence. In this simple image, turbulent ows are described as being made up by the presence of two completely di8erent regions (a) a random and structureless background which occupies most of the ow (its volume fraction :bg is very close to one). In this random background, correlations between small scales quantities at di8erent locations, such as !2 (x)!2 (x + r), are small and the overall statistics are well described by Kolmogorov K41 theory (or perhaps K62 reAnements). (b) a collection of intense vortex Alaments which occupy only a very small volume fraction (of the order of Re#−3=2 or Re#−2 . These vortex tubes correspond to regions of strong spatial correlations and are responsible of the deviations of high-order moments from Kolmogorov predictions. In that simple picture, it is for example assumed that the anomalous Re#1=2 scaling of the variance of the pressure-gradient terms is mainly due to the contributions of the intense coherent structures. 5.6.2. A Lagrangian point of view From this point of view, we consider turbulent ows as being made up by a large number N of ‘ uid particles’. This corresponds already to a discrete vision of a turbulent ow rather than a continuous one. This point of view is a statistical one in which one tries to describe the N -point pdf and from it the discrete approximation of statistical quantities. Each particle can be thought of as having the same mass for the sake of simplicity and have location xn and velocity Un . The N di8erential equations that give the evolution in time of the Lagrangian properties are written from the Navier–Stokes equations as d xin (177) = Uin ; dt d Uin (178) = Fp + Fv ; dt for n = 1; : : : ; N and where Fp and Fv stand for the pressure-gradient and the viscous forces, respectively, 1 9P Fp; i = ; Fv; i = QUi : (179) 9xi The pressure-gradient and viscous terms that enter the time evolution equation for one particle labelled n depend upon the values of all other particles, so the set of equations describes
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
61
indeed a coupled N -particle problem. From Kolmogorov theory, we know that Fv is actually a short-range force whereas Fp is a force which has both short- and long-range e8ects. The pressure-gradient force is a crucial term, particularly in a Lagrangian formalism since it connects even widely separated particles. This can be further brought out by reformulating the Poisson equation satisAed by P, Eq. (165). Using the three small scale invariants, +, ’ and C deAned in Section 5.5.1 and the relations given in Eqs. (174), we can write QP = (’ − +) = (C − +) = (180)
and by inversion of the Poisson equation, the pressure-gradient force is given by 1 1 (t; y) d y : (181) Fp = − 4( |x − y|2 This formulation reveals an interesting analogy with classical electrostatic. It is seen that the uid particle interact through Coulombian elementary forces where plays the role of the electric charge. At each point, and thus for each particle, the di8erence between the enstrophy C and the mechanical dissipation + creates a non-zero charge at that point. When the enstrophy is in excess over the mechanical dissipation, the charge is positive and this leads to attractive forces between the particles. It is thus seen that stable coherent structures must correspond to low-pressure events (or points when QP ¿ 0) and to high vorticity regions, which are indeed found in the vortex Alaments. The particle or Lagrangian point can help to clarify the di8erent meaning between the three small scale invariants. If we consider the total energy W received by a particle as the result of the work performed by the surface forces acting on that particle (nj is a normal unit vector directed outward) 9 W = Ui $ij nj d S = (Ui $ij ) dV ; (182) S V 9xj where $ij is the stress tensor 9Ui 9Ui + : $ij = −Pij +
9xj 9xj
(183)
The total work (per unit mass) is thus given by the sum of two contributions DUi W = W1 + W2 = Ui + +: (184) Dt The Arst term, W1 represents the change of kinetic energy of the particle while the second term, W2 represents the change of internal energy of the particle. We have therefore that W = ’ + + ;
(185)
(186) QP = [’ − +] :
which suggests that ’ and + do not play the same physical role: + is the rate of increase of a particle internal energy while ’ (actually the di8erence with +) appears as the rate of energy transfer between uid elements of particles. This could be related to the observed di8erent distributions of these invariants in homogeneous turbulence as discussed in Section 5.5.1.
62
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 8. Sketch of the force created by one particle on other uid elements.
It may be instructive (and perhaps also amusing) to re-derive usual Kolmogorov estimations from this Lagrangian point of view. For this purpose, let us evaluate the energy transfer through the work performed by the pressure-gradient force. We consider one particle, which has a small volume, say V0 an associated charge 0 , a characteristic velocity u0 and a typical length d0 . These ‘particles’ which are in fact small uid elements having a certain coherence are not stable. They are subject to viscous stabilizing e8ects which act with a typical time scale of the order of $v ∼ d2 = and to perturbation which act with a typical time scale of the order $p ∼ d=u. Particles are only stable when $p 6 $v , that is when the Reynolds number based on the particle length d is less than one. In other words, particles are only stable when d ∼ @. We therefore associate to each particle a characteristic lifetime, say $(d). The single particle that we are considering (and which is labelled with the index 0) creates a force on all uid elements, see Fig. 8. The force density at the distance r from the reference particle is given by 1 0 V0 4(r 2 and a small volume element dV around the point M is subject to the force f(r) = −
(187)
d F(r) = f(r) dV :
(188)
The small volume element will then have a typical velocity induced by the force created by the reference particle, say u(r) that we can write as u(r) = f(r) $0 ;
(189)
where $0 is the characteristic time during which the force will be applied on the small volume dV at location M. The work performed by the force is d P = d F(r) u(r) = T0 f(r)2 dV
(190)
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
63
and the total work, or in other words the total energy lost by the reference particle labelled 0 per unit mass becomes ∞ 1 1 1 W= P= T0 02 4 V0 4(r 2 d r : (191) 2 V0 (4() r d0 We then obtain the simple expression of the energy transfer rate for a given (and dropping now the label 0) W ∼ $(d) d2 2 :
uid particle (192)
Since the ‘charge’ is also expressed by = !2 − s2 by Eqs. (174) and (180), we can write W ∼ $(d) d2 !4 :
(193)
The simplest estimation for ! and $(d), for particles with a typical length greater than the Kolmogorov scale, are !∼
u ; d
$(d) ∼
d : u
(194)
This yields that W ∼ u3 =d and if we assume that the energy transfer for the pressure-gradient force is of the order of the characteristic energy transfer rate for the ow as a whole, thus of the order of +, we Anally retrieve Kolmogorov scaling that is u3 ∼ + : (195) d The simple two-fold description of turbulence put forward in the preceding subsection and the derivation of the energy lost by work of the pressure-gradient force can be used to suggest modelling ideas for this pressure-gradient term. If we leave out the (small) region involving the coherent structures and if we consider only the structureless background, then with Eq. (181) and using the assumption that the charge density (x) is nearly uncorrelated in space (x)(x + r)1, we obtain that the pressure-gradient Fp should be close to a Gaussian random variable as the consequence of the Central Limit Theorem. Using a di8erent terminology, we can say that in the random background, pairwise interactions between uid particles are weak and are similar to ‘grazing collisions’ of particle physics for which the Bolztmann equation is sometimes replaced by a Fokker–Planck equation. These two points, loss of energy through the transfer towards all other uid particles and a near Gaussian form when coherent structures are ignored, suggest to consider the simple mechanical models often retained for the motion of a charged particle in classical electrodynamics m
d2 x dx + < + kx = Fext : dt 2 dt
(196)
In classical electrodynamics, a charged particle which accelerates (in a general sense this means its velocity is not constant) creates an electromagnetic Aeld and therefore a force Aeld that acts on all other charged particles. This force Aeld corresponds to a transfer of energy from the charged particle that we are considering to the others. In that sense, it is said that an accelerating charged particle ‘radiates’ energy: its own energy E decreases due to the transfer written
64
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
as −d E= d t to all other particles. In the simple mechanical model, the friction coe@cient < in Eq. (196) is meant to represent the radiative damping coe@cient, thus 1 dE <=− : (197) E dt We can use similar ideas for the pressure-gradient term building on the above-mentioned analogy with electrodynamics. In an homogeneous ow, where the mean velocity U is zero, this analogy suggests to model Fp by Fp = −
(198)
where the friction coe@cient is expressed as W <∼ 2 ; (199) u with W the energy transfer and u2 the particle kinetic energy. Using the classical estimations in the K41 picture, W ∼ +, we obtain 1 + <∼ 2 ∼ ; (200) u TL where TL is the Lagrangian integral time scale deAned in Section 5.2 (see also Eq. (134)). Furthermore, we can take the uctuating term fp as a Gaussian random term in the structureless background as proposed above. We then obtain a model equation which has the form of a Langevin equation, and generalising to non-zero mean velocities, this model can be written as 1 Fp ∼ − (U − U ) + G ; (201) TL with G being a Gaussian white noise. These ideas are of course just simple local models and should just be regarded as crude representations. Nevertheless, they are not necessarily without any connections to the physics of turbulence and they still retain some of the essential aspects of Kolmogorov theory. It will be seen in the following sections that they form the basis of present pdf models. 5.7. Closing remarks A great body of work has accumulated over the years on turbulence, but the subject has retained its mysteries. At that point, it may be useful to repeat that di8erent objectives can actually be pursued when dealing with turbulent ows. To somewhat simplify the picture, we can distinguish between a theoretical point of view and a modelling point of view. From the theoretical point of view, one looks for universal features and the concern is mainly on small scales, scaling exponents, intermittency and high-order statistics. The central issue is then to remedy to the deAciencies of Kolmogorov theories. From a modelling or ‘engineering’ point of view, the emphasis is more on mixing and transport phenomena which are mainly governed by the large energy-containing scales. The two points of view are not completely separated, and, as explained in the introduction of the paper, the purpose of the present work is actually to And a middle-road approach and
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
65
a compromise between the two. An example may be the Arst and simple modelling steps presented in Section 5.6.2. It is believed that theoretical improvements or new results, such as the determination of small scale statistics, can help even application-oriented models. However, from the modelling point of view, it is clear that high-order scaling exponents is an issue which is not of Arst importance. The most important aspects are well identiAed in Kolmogorov theories and in the description of the energy cascade. Even the Arst K41 theory does not describe a system in thermodynamical equilibrium, but rather a system maintained out of equilibrium by a constant ux. A Arst-importance issue is to determine the energy transfer rate from the large to the dissipative scales. The uctuations of this variable are important for practical concerns as well as for theoretical aspects (see the discussion of recent results [45] in Section 5.5.1). The determination of the energy transfer rate is a very di@cult point since the large scales are not isotropic and are not universal. It is thus seen that the turbulence problem is not, in that respect, only due to the large number of degrees of freedom but to the strong interactions between them, in particular between the large scales. Another main di@culty is that, at the moment, no small parameter has been identiAed from which successive approximations could be rigorously derived. Depending on the point of view, small parameters do appear but they only concern partial aspects of the whole problem. For example, by adopting a Lagrangian formalism, we have seen that there exists a separation of scale (of characteristic time scales) between uid particle velocity and acceleration. This suggests already a Lagrangian point of view and the small parameter Ta =TL (or even $@ =Qt). Yet, deriving a local or one-particle model remains di@cult since one has to model (for instance as in Section 5.6.2) the interactions of the large scales. These models are then uncontrolled approximations, or, in other words, guesses. On the other hand, one should not overlook their practical interest as we will see in the next sections. The discussion related to large scale direct interactions is taken up again in Section 9.3.2. 6. One-point pdf models in single-phase turbulence The mathematical tool of stochastic processes (particularly, di8usion processes) and the general modelling ideas that have been developed so far can now be applied for turbulent ow modelling issues. The objective of the present paper is to formulate a pdf approach for two-phase ows, but it is useful to consider only single-phase turbulent ows as a Arst step before addressing the two-phase ow problem. Presentation of the pdf approach for single-phase turbulent reactive ows can be found in previous works, for example [6,10,8,9] or [54]. Therefore, we do not try to give a comprehensive presentation, but rather choose to insist on certain aspects of the approach. In particular, the emphasis in this section is put on (i) clarifying the purpose of a pdf description in turbulence, with respect to other particle methods and with respect to classical approaches in turbulence, (ii) detailing the formalism, especially the notions of Lagrangian and Eulerian pdfs, (iii) explaining the reasoning behind the choice of a Lagrangian description, of a precise state vector and of di8usion processes as models, (iv) outlining the physical content of present state-of-the-art stochastic models and their range of assumed validity.
66
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Some of these aspects are not always covered in the previously cited works, or may complement them. Details on the formalism can already be found in Pope [6] but are presented here from a slightly di8erent point of view. The general formalism and the stochastic models used for the two-phase ow modelling issue are actually extensions of those given in this section. In the following, · will denote the classical Eulerian averaging operator (whose exact mathematical deAnition will be properly given in Section 6.4). 6.1. Motivation and basic ideas As explained in Section 5, a turbulent ow displays far too many degrees of freedom: the range of scales can cover several decades, i.e. L=@ and T=$@ 1, and is therefore too extended to allow a direct calculation even with computers of the foreseeable future. Thus, although the basic equations are deterministic, the Arst modelling step consists in adopting a statistical point of view and in regarding the solutions of the basic equations as being random or stochastic processes. This is actually a classical step in statistical physics where probabilistic arguments are used in (deterministic) systems involving a very large number of degrees of freedom. This step goes hand in hand with the fact that we are not interested in the full description of the ow in time and space but rather in some limited information about some characteristics of the Aelds. In the language of mathematics, we would say that we are interested in a weak approximation (see Section 2.1.2). In the language of statistical physics, we are interested in a reduced or contracted description of continuous Aelds. 6.2. Coarse-grained description and stochastic modelling At that point, it is worth recalling that the basic equations, the Navier–Stokes equations, have precisely been obtained through such a procedure, e.g. Balescu [22] and this is an outstanding example of the success of a reduced description. Their derivation follows the steps: molecular equations : microscopic description ↓ Boltzmann equation : kinetic description ↓ Navier–Stokes equations : macroscopic description : In this procedure, the microscopic equations of motion are not treated exactly but are bluntly replaced by a probabilistic model (where the assumption of molecular chaos is input). Once this is done, the macroscopic equations of motion are derived from the kinetic equation. The downward orientation of the arrows indicates a contraction of the information and an evolution towards less detailed descriptions of the system. The kinetic equation approach is called a mesoscopic description. In turbulence modelling, the need to limit ourselves to a reduced set of degrees of freedom shows that we are faced with a similar task. The di8erence lies in the starting point which is now the very hydrodynamical equations. What is regarded as ‘macroscopic’ in the classical kinetic approach is now taken as being the ‘microscopic’ level. The procedure can
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
67
be sketched as Navier–Stokes equations : microscopic description ↓ ‘model’ equations : mesoscopic description ↓ mean Reynolds equations : macroscopic description : Here, the adjectives of microscopic and macroscopic lose their meaning of small and large from a geometrical point of view and refer more to Ane and coarse descriptions in terms of the number of degrees of freedom involved. The idea of the present approach is to introduce a model at some intermediate level between the exact Navier–Stokes equations and those for the mean moments in which we are interested. The construction of a reduced state vector can be achieved by a coarse-graining procedure where the system is described on a ‘large enough scale’ to eliminate some degrees of freedom. Information is therefore lost and this lack of complete knowledge will be re ected by the use of a stochastic description for the remaining degrees of freedom. Resorting to a statistical description is common to all turbulence models. However, the key point is that we will not try to write directly a model in terms of macroscopic variables but, following the above sketch, we will try to introduce the model at an upstream level, or, using the same terminology as above, at a mesoscopic level. In other words, the aim is to have a pdf description through the Aeld of probabilistic density functions, and not simply the knowledge of a limited number of moments at each point (usually not more than two). 6.3. Relations to classical approaches First of all, present pdf models should not be confused with other particle methods, such as Vortex Methods or S.P.H. (smoothed particle hydrodynamics) in their basic forms. Indeed, in random vortex methods one interprets the Navier–Stokes equations (written in the vorticity form) as pdf equations to be solved by a particle method. This corresponds exactly to the probabilistic interpretation of PDEs (here Navier–Stokes) mentioned in Section 2 and the particles used in this method are ‘computational particles’ whose distributions give the solutions of the Navier– Stokes equations. As such, this is equivalent to a direct numerical simulation with particles. It is also useful to contrast Lagrangian stochastic models with classical turbulence approaches. The classical approach starts by applying some averaging or Altering operator to the exact equations. We obtain exact but unclosed ‘mean’ equations in which closure relations are then introduced. Closed ‘mean’ equations result. For example, in the moment approach one writes the mean Navier–Stokes equations (we assume constant density ows for the sake of simplicity) 9ui uj 9Ui 1 9P 9Ui =− + QUi − ; + Uj 9t 9xj 9xi 9xj
(202)
in which the unknown Reynolds stress tensor ui uj has to be expressed. The mean momentum equation can be closed by assuming a constitutive law for the Reynolds stresses (this is the
68
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
spirit of the k–+ model) or by writing their own exact equation 9ui uj 9ui uj 9ui uj uk =− + Uk 9t 9xk 9xk 9Uj 9Ui Pij −ui uk − uj uk 9xk 9xk
1 9p 1 9p − Hij − ui uj 9xj 9xi ’ij + ui Quj + uj Qui :
(203)
The closure problem has only been shifted to the rate of change of ui uj and one has now to come up with closures for the triple correlation ui uj uk , the pressure rate of strain correlation Hij and the work of the viscous forces ’ij . The latter can be decomposed as
9uj 92 ui uj 9ui − 2
; (204) ’ij =
9xl 9xl 9xl2 where the last term represents twice the pseudo-dissipation tensor written as +ij . Much of the current research in second-order modelling so far has focused on the closure of the Hij terms (whether or not this is fully justiAed). Yet, closure expressions for the di8erent Hij are not only di@cult (since limited information is available), but as a result of the approach itself, it is also di@cult to be sure that they are coherent among themselves. This is called the realizability problem. More than the particular details of certain models, it is important to underline the steps that are taken in this classical approach: Navier–Stokes equations ↓ open or unclosed mean Reynolds equations → introduction of a model: closed mean equations .
Therefore closure is attempted directly at the macroscopic level, and when it is performed, available information is of course strictly limited to the very macroscopic variables that have been explicitly retained in the second step of the procedure. Compared to this, in the pdf approach, the introduction of a model is made at an ‘upstream’ level, as indicated in the sketch in Section 6.2, where far more information is still available (since we model a probability density function) and has not yet been eliminated. As a result, pdf approaches are particularly attractive if (i) detailed information is needed to address a problem, at least more than the mere knowledge of a few moments, (ii) it is too di@cult to close directly at the macroscopic level through constitutive relations between macroscopic variables. The Arst item is related to the fact that, depending upon the problem involved, one may be interested in more than one or two mean values. A clear example of the second item will be given in Section 6.5.3 for reactive ows. The di8erence can be further illustrated by the
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
69
following representation of the pdf approach Navier–Stokes equations ↓ open or unclosed pdf equations → introduction of a model: closed pdf equations ↓
closed mean equations .
where as above, a horizontal arrow stands for the introduction of a model closure at the same level of description, while a vertical arrow means that we move from one level of description to a coarser one. 6.4. Probabilistic description of continuous ?elds We have referred to classical kinetic theory and to the Boltzmann approach in order to bring out the ideas and the purpose of present pdf models for turbulent ows. However, the probabilistic description needed in the pdf approach is more complex than in the Boltzmann approach and it requires that the notions of particles and Aelds, or Lagrangian and Eulerian points of view, be clearly speciAed. In classical kinetic theory, there is a clear di8erence between the meanings and the way with which the notions of particles and of Aelds enter the theory. We are dealing with an ensemble of ‘real’ particles and the system on which we apply a probabilistic description is discrete. It is therefore not surprising to use a particle point of view in the probabilistic description and to handle a Lagrangian probability density function or distribution function. There is no reason in that context to use a di8erent point of view. On the other hand, the resulting local mean values or moments derived from the pdf (or kinetic) description are Aelds: the solutions of the Navier–Stokes equations. Thus, there is a clear di8erence between the two notions of particles and of Aelds in classical kinetic theory. The microscopic level, which is made up by discrete particles, is described from a Lagrangian point of view, and the macroscopic level, which is made up by the Aelds which are solutions of the Navier–Stokes equations, is described from an Eulerian point of view [22]. In the present pdf approach for turbulent ows, there is no such clear-cut separation between the two notions. We do not give a statistical description of a discrete system but instead a probabilistic description of Aelds using stochastic processes. This can create some confusion with respect to the notion of Lagrangian and Eulerian descriptions, or particles and Aelds. Indeed, in this case, the system considered (the microscopic level, now represented by the solutions of the Navier–Stokes equations) can be described either from a Lagrangian or from an Eulerian point of view. The latter appears natural in the context of continuum mechanics, but the former can also be adopted to characterize the system. As a result, the probabilistic description can either be performed in an Eulerian or in a Lagrangian framework. Yet, the outcome of the whole approach, which are the di8erent mean values or the local moments of the variables of interest, will still be Aelds. In other words, the resulting macroscopic quantities will be expressed in an Eulerian frame, but the road leading to them can be either a Lagrangian or an Eulerian one.
70
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
The purpose of the following subsections is to deAne the Eulerian and Lagrangian pdfs and the way they are related. 6.4.1. Dimension of the state vector The following deAnitions and relations between probabilistic quantities will be developed with a certain expression of the state vector. For this we slightly anticipate in Section 6.6 where the choice will be discussed, and we already use a state vector. That state vector corresponds to a one-particle description and to a one-particle state vector which includes particle location, velocity and a number of scalars denoted by in physical space or in sample space, see Section 3 Z = (y; V; ) :
(205)
At present, the meaning of the variables can be left undeAned, standing for any particle properties, and the state vector Z remains therefore fairly general. 6.4.2. Eulerian and Lagrangian descriptions First of all, the exact one-point pdf equation satisAed by pL (t; y; V; ) can be derived by writing the Navier–Stokes equations and a scalar equation in a Lagrangian formulation. The probability to And the reduced system—a ;uid particle—in the state [y; y + d y], [V; V + d V] and [ ; + d ] is given by pL (t; y; V; ) d y d V d ;
(206)
where the superscript L is introduced to emphasize that we are dealing with a Lagrangian formulation. The two possible points of view (Eulerian or Lagrangian) di8er essentially by the choice of the independent variables (or parameters). It is of common knowledge that in the Lagrangian description, attention is focused on ‘labelled’ uid particles as they move through the ow (initial position and time as parameters, position, velocity and scalars as variables) whereas in the Eulerian approach (Aeld description), all ow properties are by deAnition monitored as functions of time and a Axed point in space (time and space coordinates as parameters, velocity and scalars as variables). Therefore, we deAne the Eulerian pdf pE (t; x; V; ) where pE (t; x; V; ) d V d
(207)
is the probability to observe in the ow Aeld a state in the range [V; V + d V] and [ ; + d ] at time t and position x. By doing so, we have deAned a Aeld of distribution function since for each point (t; x) of the time-physical space, a distribution function of velocities and scalars has been associated. To emphasize the distinction between the two descriptions, a semi-column is introduced to separate the parameters from the variables, i.e. pE (t; x; V; ) and pL (t; y; V; ). Note that we also distinguish between phase space and physical space: (y; V; ) ↔ (x; U; ) for the Lagrangian description and (V; ) ↔ (U; ) for the Eulerian one. To complete this formalism, we need to deAne physically the notion of ;uid particles. A uid particle is a small element of uid whose characteristic length scale is smaller than the Kolmogorov length scale, Eq. (135) (which is the smallest length scale of uid motion) but much larger than the molecular mean free path. The uid particle (which can be deformed) has a mass m(t; ), a volume V (t) where (t; ) = m=V and a velocity U(t) equal to the value of
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
71
the Eulerian velocity Aeld at the location of the uid particle, U(t) = U(t; x(t)). This notion is in fact a direct extension of the particle (point) notion in classical mechanics. In the case of particles of constant volume (the arbitrary choice that we make from now on), the ;uid particle object is fully deAned by the set of following variables: (m; y; V; ). Therefore, for a complete description, it is convenient to introduce a mass density function F L (t; y; V; ) where F L (t; y; V; ) d y d V d ;
(208)
represents the probable mass of uid particles contained in an element of volume d y d V d in phase space. The mass density function is consequently normalized by the total mass of uid M (which is chosen to be constant in time for the sake of simplicity) M = F L (t; y; V; ) d y d V d : (209) The Lagrangian mass density function is given by F L (t; y; V; ) = M pL (t; y; V; ) :
(210)
6.4.3. Relations between Eulerian and Lagrangian pdfs We have previously introduced the notion of Eulerian and Lagrangian pdfs and it has been explained why in the case of compressible or reactive ows, the proper pdf to be considered is the mass density function F L and not the pdf itself pL (which is of course the right choice in the incompressible case since all uid particles have a constant mass for a given volume). Attention is now focused on the possible relations between the two points of view. Indeed in order to derive partial di8erential equations veriAed by the mean moments, one has to be able to deAne the proper Eulerian quantity to construct an operator which gives the expected values (moments). The techniques presented in Section 2 are ‘Lagrangian tools’ and moment equations which are Aeld equations can only be derived by means of ‘Eulerian tools’. The question to be answered is: how are Eulerian and Lagrangian quantities related? Following Balescu [22], the correspondence between the Eulerian and the Lagrangian descriptions is given by E L F (t; x; V; ) = F (t; y = x; V; ) = F L (t; y; V; ) (x − y) d y ; (211) where F E is the Eulerian mass density function. This relation simply expresses the fact that, when in phase space y = x, there is equivalence between the two points of view since we consider the same event (in the incompressible case this relation is veriAed by pE (t; x; V) and pL (t; y; V)). By integration of the previous equality over phase space (V; ) and physical space, we can write (212) M = F E (t; x; V; ) d x d V d : This constraint imposes that the integral of F E over phase space (V; ) is the expected density at (t; x) (the probable mass of particles in a given state per unit volume). The expected density,
72
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
denoted (t; x), is of course deAned by (t; x) = ( ) pE (t; x; V; ) d V d ;
(213)
(the integral of pE over the (V; ) phase space equals one since pE is a pdf) and since the integral of the expected density (t; x) over the whole physical space is of course the total mass of uid, the Eulerian mass density function is deAned by F E (t; x; V; ) = ( ) pE (t; x; V; ) :
(214)
The relation between both mass density functions has now been deAned, Eq. (211), and the expression of the Eulerian mass density function is known, Eq. (214). The remaining task is to And the relations between the pdfs. Intuitively, one might think that the Eulerian pdf equals the conditional (conditioned by the position) Lagrangian pdf. Let us investigate this matter. Integration of both mass density functions over the (V; ) phase space gives F L (t; x; V; ) d V d = M pL (t; x) ;
F E (t; x; V; ) d V d = (t; x) ;
(215)
which gives pL (t; x) = (t; x)=M . The conditional expectation pL (t; V; | x) can then be evaluated as follows (using the previous results): pL (t; V; | x) =
pL (t; x; V; ) F L (t; x; V; ) ( ) = = pE (t; x; V; ) : L p (t; x) (t; x) (t; x)
(216)
Therefore, in a compressible ow (and also in a turbulent reactive ow) the Lagrangian pdf conditioned by the position is not the Eulerian pdf but the Favre pdf. Finally, the relation between the mass density function and the Lagrangian transitional pdf can be worked out from the deAnition of the Lagrangian pdf using the transitional pdf L p (t; x; V; ) = pL (t; x; V; | t0 ; x0 ; V0 ; 0 ) pL (t; x0 ; V0 ; 0 ) d x0 d V0 d 0 : (217) By multiplying each side by the total mass M and using the identity given by Eq. (211), we obtain E F (t; x; V; ) = pL (t; x; V; | t0 ; x0 ; V0 ; 0 ) F E (t; x0 ; V0 ; 0 ) d x0 d V0 d 0 : (218) This relation is of extreme importance: it shows that the mass density function is ‘propagated’ by the transitional pdf, or in the language of statistical physics, the transitional pdf is the propagator of an information which is the mass density function. Before, we carry on, let us rewrite the results derived in this section in the particular case of an incompressible ow and from now on, only this type of ows will be studied. When the ow is incompressible, uid particles have a constant mass (density is constant) and all the information is contained in the pdfs. Eq. (211) shows then that pE and pL are related by
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
(with = M= V where V is the volume of the physical space occupied by the ow) 1 E L p (t; x; V) = p (t; y = x; V) = pL (t; y; V) (x − y) d y : V
73
(219)
The probability to And a uid particle at a given position x is associated to the marginal pL (t; x) = 1= V. This result is physically sound since in the case of incompressible ows mass is equally distributed in space and all events are equiprobable. The Eulerian pdf is, in the incompressible case, the Lagrangian pdf conditioned by position, cf. Eq. (216), pL (t; V | x) =
pL (t; x; V) = pE (t; x; V) : pL (t; x)
At last, for the incompressible case, Eq. (218) is veriAed by pE E p (t; x; V) = pL (t; x; V | t0 ; x0 ; V0 ) pE (t; x0 ; V0 ) d x0 d V0 ;
(220)
(221)
which means, as mentioned previously, that the transitional pdf is the propagator of an Eulerian information which is, in this particular case, the Eulerian pdf. 6.4.4. Discrete representation and weak approximation The di8erent mdfs and pdfs have been presented so far in a continuous framework, and as being continuous densities or measures. In actual simulations, these measures are approximated by discrete measures using a Anite number of particles, and practical implementations require to manipulate these discrete mdfs. Nevertheless, it appears preferable to separate the handling of these approximate densities from the deAnitions of the actual densities (the Lagrangian and Eulerian mdfs and pdfs), so as to keep the presentation of the key notions as clear as possible. In a practical simulation, a Anite number of particles are followed in time. Each of these particles have a particular value of its attached variables, namely x(t); U(t) and (t). The discrete approximation of the Lagrangian mass density function is therefore given by N FNL (t; y; V; ) = Qm (y − xi (t)) ⊗ (V − Ui (t)) ⊗ ( − i (t)) ;
(222)
i=1
using N particles with equal mass Qm. When particles have di8erent masses, the discrete Lagrangian mdf has the form FNL (t; y; V; ) =
N i=1
mi (y − xi (t)) ⊗ (V − Ui (t)) ⊗ ( − i (t)) :
(223)
The total expected mass in the volume occupied by the ow is then M = Ni=1 mi or, in the case of equal particle mass M = N × Qm. The approximation of a mean Eulerian quantity (Favre averaged) at a given point in space x and at time t, H˜ (t; x) can be obtained from the above discrete expression of FNL . Indeed, the density-weighted integral is (224) (t; x)H˜ (t; x) = () H (U(t; x); (t; x)) = H (V; ) F E (t; x; V; ) d V d :
74
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Using Eq. (211) where the delta function in space is replaced by (y − x) 1=Vx when y ∈ Vx , Vx being a small volume around point x, we obtain (t; x)H˜ (t; x)
N
x 1 mi H (Ui (t); i (t)) : Vx
(225)
i=1
With the local (discrete) mean density being given by Nx i m (t; x) i=1 ; Vx
(226)
it is seen that the discrete approximation, through the ensemble average over the Nx particles present in the small volume Vx around x, approximates the Favre-average mean quantity H˜ Nx i m H (Ui (t); i (t)) ˜ H HN = i=1 Nx : (227) i i=1 m In other words, the Favre-average mean value of any function of the local Eulerian velocity and scalars, is obtained by a Monte Carlo calculation from the local number of particles. The procedure is valid in the general non-stationary and non-homogeneous case, but is actually based on a local homogeneity hypothesis. A small volume Vx is introduced around point x, and the Nx particles which are located in this small volume are regarded as independent realizations and samples of the mdf at point x. This amounts to assuming spatial homogeneity within the small volume Vx . Convergence of the discrete approximation is ensured by the Central Limit Theorem which shows that there exists a constant C such that HN = H˜
and
(H˜ − HN )2 6
C : Nx
(228)
It is therefore seen that the discrete distribution FLN does not tend towards FL in a strong sense but in a weak sense, or to be more precise in distribution (see Section 2.1.2), since in fact it is the mean value of functions of the stochastic process Z that converges as Nx → +∞ where Nx stands for the simulated number of samples of Z HN = H (ZN ) → H (Z) : N →∞
(229)
This notion of weak convergence and of weak approximation is fully consistent with the overall aim of the pdf approach where, as explained in Section 6.1, the interest is to obtain information on various statistics of the ow. In practice, the challenge is to reach a satisfactory compromise between spatial errors (governed by the size of Vx ), time discretization errors (governed by the time step Qt used to integrate the trajectories of the process, see Section 2.10) and statistical errors (governed by Nx1=2 ) for a tractable total number of particles. This constitutes the important Aeld of practical Monte Carlo methods, see Kalos and Whitlock [55] for example and specialized articles [6,54,56].
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
75
6.5. Choice of the pdf description 6.5.1. The Lagrangian stochastic point of view From the previous section, it appears that the central notion is the transitional Lagrangian pdf from which Lagrangian and Eulerian mass density functions are derived. Consequently, the probabilistic description is best performed using a Lagrangian point of view that will be followed from here onwards. From that point of view, the Navier–Stokes equations are regarded as expressing directly the time-evolution equations of a large number of uid particles. These exact equations are replaced by models which (hopefully) produce the same statistics as those of real uid particles. Following the reasoning of Sections 6.1 and 6.2, these models are stochastic models. In turn, these stochastic models can be developed and expressed from two points of view, either a pdf point of view or a trajectory point of view, see Section 2. The second important choice that is made is to use mainly the trajectory point of view. In a way, we can say that we ‘merge’ two notions into one, the particle or Lagrangian description in turbulence on the one hand and the trajectory point of view for the expression of stochastic processes on the other hand. A somewhat condense presentation of these steps would be to say that the exact instantaneous equations are replaced by modelled but still instantaneous ones. 6.5.2. PDF equation in single-phase turbulence Let us recall the original problem, i.e. which is the formulation of the exact one-point pdf equation satisAed by pL (t; x; V; | t0 ; x0 ; V0 ; 0 ). We start by writing the Navier–Stokes equations and a scalar equation in a Lagrangian formulation (for a given uid particle) d xi+ = Ui+ d t; + 1 9P d Ui+ = − + QUi dt ; 9xi + + d + l = >Q l d t + Sl ( ) d t :
(230)
Note that the subscript ‘+’ has been used to distinguish between the exact equations and the modelled ones. By applying for example the techniques presented in Section 2 (other techniques can also be used), one obtains (where pL+ stands for the transitional pdf associated to the exact trajectories)
9pL+ 9pL+ 1 9P 9 9 L+ − − =− Z=z p [ QUi | Z = z pL+ ] + Vi 9t 9yi 9Vi 9xi 9Vi −
9 9 [>Q l | Z = z pL+ ] − [Sl ( ) pL+ ] : 9 l 9 l
(231)
In this equation, A|Z = z represents the average value of A conditioned on the values of the vector state Z = z, cf. Section 3. Then, Eqs. (217) and (218) show that the same equation is satisAed by the Lagrangian pdf and by the Eulerian mass density function. From the latter, mean moment equations can be expressed (see Section 6.7).
76
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
6.5.3. Interest of the pdf approach The two main features of Eq. (231) are that convection and the scalar source term appear in closed form. No modelling assumption is required to get a closed form for these terms. First of all, treating convection without approximation means that whatever expression coming out of the convection term, in any moment equation, does not need to be modelled. Therefore, the closure di@culties that we mentioned with the classical moment approach, due to ui uj or ui uj uk are not met in our case. This is already very attractive since, from our discussion of statistical closures and our presentation of modelling principles, one can expect that replacing correlations arising from the convection term (and Reynolds stresses are nothing but that) by accurate models is di@cult. The second feature is even more noteworthy. However complicated and non-linear the source term may be, it is in closed form in the present approach. On the contrary, in the classical approach to turbulence modelling (the moment approach), one has to express S() directly in terms of available information, which is limited to () and ()2 for complex functions S. Such a direct closure at the macroscopic level is hopeless, unless speciAc and restrictive assumptions are made (fast chemistry, etc.) and this is a very clear example of the second interest for pdf descriptions as emphasized at the end of Section 6.3. Indeed, the calculation of the mean source term is handled correctly since it is given by S() = S( ) pE (t; x; ) d ; (232) where pE (t; x; ) d is now known. This success explains the great interest of present models for problems involving similar source terms such as in combustion. These advantages are easily explained by the present Lagrangian point of view. Following particles ensures that convection is treated without approximation. Furthermore, since we keep track of instantaneous values, any local source term such as S() is also treated without assumption. This is easily seen by the Monte Carlo evaluation of the mean source term using the Nx discrete particles which are located in a small volume around the point x at which the integral is evaluated as Nx 1 S() S()Nx = S( (i) ) : (233) Nx i=1
On the other hand, the terms involving the pressure-gradient and the molecular transport coe@cients are not in closed form. This is not surprising from the particle equation of motion: these terms represent the instantaneous forces or actions exerted by the other particles (other elements of uid). This expresses the BBGKY hierarchy of unclosed pdf equations explained in Section 3. In order to develop a practical model in a pdf framework, the state vector and its dimension must be chosen. If we adopt the picture of a turbulent ow as a N -particle problem (the continuum is retrieved by taking N → ∞), we must choose the values of N and of p, see Section 3. One-point models can only provide one-point information (all mean moments) but no space correlations at a given time, since these correlations are in fact contained in a two-point description. However, at present, two-particle models are not as developed as one-particle ones, especially in the general case of non-homogeneous non-stationary ows. We therefore consider
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
77
only one-point or one-particle models in the following (in the notations of Section 3, we have s = 1). The remaining choice is the one-particle state vector. 6.6. Present one-point models Once the pdf stage is set, remains the question of the proper choice of the state variables that will be kept in the state vector and those that will be eliminated in turbulence modelling. From the Kolmogorov’s hypothesis described in Section 5.3, uid particle accelerations are close to white-noise processes. Thus, the uid particle acceleration is a good candidate to be replaced by a model and there is some hope to derive a one-point model by restricting the state vector to particle locations and velocities. This justiAes the choice of the one-particle state vector, that we have already used namely Z = (y; V; ). In this section, models for uid particle location and velocity are presented. Models for scalar variables are left out and can be found in Pope [6], Fox [9] or Dopazo [8]. This could be surprising since one of the main interests of pdf models, for single-phase turbulent ows, is precisely the treatment of reactive scalars. However, the aim of the present paper is the presentation of pdf models for two-phase turbulent ows where particle velocity stochastic models have similar interest. The stochastic models described here will be used as starting points in later sections. 6.6.1. Stochastic model for ;uid particle velocity A model is now derived for pL (t; x; V). From the instantaneous particle equation of motion Eq. (230), the two forces acting on a uid particle, the pressure-gradient force and the viscous force, have to be expressed as functions of local variables (one-point closure). A global model can be proposed for the sum of these two forces and analysed with respect to classical second-order modelling [7,10]. Another approach is to build a model in two steps, by modelling separately the viscous term and the pressure-gradient term, relying furthermore on arguments from Statistical Physics rather than on macroscopic mean equations. The starting point is to reformulate the exact one-point pdf equation, Eq. (231). We decompose the pressure P into a mean component P and a uctuating one p , and it can be shown that the viscous term can be re-expressed as the sum of two contributions [6,11]
9pL+ 9pL+ 1 9P 9pL+ 9 1 9p L+ = + Z=z p + Vi 9t 9yi 9xi 9Vi 9Vi 9xi +
92 [ pL+ ] 92 − [+ij |Z = zpL+ ] ; 9Vi 9Vj 9yi2
(234)
where +ij is the pseudo-dissipation tensor deAned in Section 6.3. By neglecting the third term involving molecular viscosity at high Reynolds numbers and by assuming an isotropic form for +ij using only the unconditional dissipation of the turbulent kinetic energy +, we obtain the working form of the one-point pdf equation
92 1 9p L 9p L 9 1 9p 1 9P 9pL L L = + Z=z p − +p : (235) + Vi 9t 9yi 9xi 9Vi 9Vi 9xi 9Vi 2 3
78
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
The superscript L+ has been replaced simply by L since we are already dealing with a model equation. It is interesting to see that the only remaining trace of the molecular viscosity is through the Anite and non-zero mean dissipation rate, +, in line with Kolmogorov theory, and that we end up with a second-order term in the pdf equation, however with a negative sign. A complete pdf model is worked out by putting forward a model for the uctuating pressure-gradient force as a function of the particle velocity and local mean quantities. A recent proposal is to resort to Onsager’s hypotheses [11] and to write a Langevin equation for the e8ect of the uctuating pressure-gradient (an external force in the one-point pdf description since this is a force created by all the other uid particles in the ow). Apart from showing that present pdf models can be more than just rough guesses and indications of some interesting links with other physical Aelds, this connection is also useful to reveal the interests and the inherent limitations of present closure relations. The real issue is to write a model to describe what is essentially a non-equilibrium statistical problem. Onsager’s hypotheses are valid mainly in the near-equilibrium domain, and their use reveal that present models cannot claim to fully describe the complete problem, and particularly that far-from equilibrium phenomena such as coherent structures may not be captured with present closure proposals. Extensions of the stochastic models or perhaps multi-point pdfs may be needed to include explicitly the statistical signature of these structures. The Langevin model for the uctuating pressure-gradient force is combined with the deterministic mean pressure-gradient and the viscous term to build a complete Langevin model for uid particle velocities. The derivation is also an example of the equivalence and the interest of the two points of view (trajectory and pdf) in stochastic modelling, since both frameworks are used in the course of the construction. Details on the modelling steps can be found in other works [54,11] and only the resulting global model is presented here. The complete model consists in retaining the instantaneous positions and velocities in the state vector Z = (x; U) and in using a general di8usion process to simulate its time rate of change d xi = Ui d t d Ui = −
(236a)
1 9P d t + Di d t + C0 + d Wi : 9xi
(236b)
The drift vector is given by the expression Di (x; U) = −
Ui − Ui + Gija (Uj − Uj ) ; TL
where TL stands for a time scale given by 1 k TL = 1 3 : ( 2 + 4 C0 ) +
(237)
(238)
It is thus seen that the uid velocity equation has the form of a Langevin equation. The drift coe@cient has actually a very simple form, involving only return-to-equilibrium terms (where the mean velocity at particle locations acts as the local equilibrium value), and it seems di@cult, apart from forgetting turbulence altogether, to obtain simpler stochastic models. The resulting model has indeed limitations but contains already interesting turbulent features [54,11]. When
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
79
the matrix Gija is put to zero, which yields the simplest model, the time scale TL represents also the time scale of velocity correlations. This is how the model appears from the trajectory point of view. In the pdf point of view, it consists in saying that the pdf equation is modelled by a Fokker–Planck equation 1 9P 9pL 1 92 9p L 9p L 9 = − [Di (x; V)pL ] + [C0 +pL ] : + Vi 9t 9xi 9xi 9Vi 9Vi 2 9Vi 2
(239)
6.6.2. Stochastic model for the dissipation rate To obtain a self-su@cient model, information about +, the mean dissipation rate of turbulent energy at particle location, is needed. + is given by
9ui 9ui + =
: (240) 9xl 9xl In a one-particle approach, instantaneous gradients cannot be derived from available information. They remain external to the description and must be provided by another source of information. Although just + is strictly needed in the particle velocity equations written above, a model for the instantaneous values of + along the particle trajectories is developed. The present model is constructed, not directly in terms of +, but rather in terms of the frequency rate !=+=k whose mean is the inverse of the time scale. The derivation of the model is explained in detail in other papers [57,58] and is based on Kolmogorov’s third hypothesis, namely that in homogeneous turbulence ln(+) is log-normally distributed. In non-homogeneous turbulence, the complete form of the model is [57] d ! = −!!S! d t + !2 h d t
! ! ! − ln d t + ! 2CK !2 d W ; − !!CK ln ! ! !
(241)
where 2 is the variance of ln(!= !) and is taken here as a constant. The drift term involving h is added to represent the mechanism through which laminar particles become turbulent and plays a role at the edges of free shear ows. The second drift term involving S! represents the normalized decay rate of the mean frequency and is modelled by reference to the standard equation of + S! = (C+2 − 1) − (C+1 − 1)
P
+
where P = −ui uj
9Ui : 9xj
(242)
The di8erent parameters, CK ; C+1 and C+2 are constants of the model. Apart from providing the needed value + at each point, the introduction of instantaneous particle values of ! allows external and internal intermittencies to be simulated (see [59]). Another model for the simulation of the instantaneous values of ! can be found in [60]. With the stochastic model for !, the complete vector state is now Z = (x; U; !) and the general pdf follows a Fokker–Planck equation in the corresponding sample space.
80
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
6.7. Mean ?eld equations The mean equations that correspond to the present Lagrangian stochastic models can be explicitly written. Once again, there are two possible ways to derive them. The Arst one is by writing the pdf equation and by integrating over all sample space after multiplication by the proper variable and the second one is a derivation directly from the particle stochastic di8erential equations (trajectory point of view). 6.7.1. pdf point of view The procedure developed here is identical to the one which is used in Kinetic Theory to obtain the Navier–Stokes equations, cf. e.g. [22,21], but at another scale as it was explained at the beginning of this section. The transport equation for pE (the kinetic equation in Kinetic Theory) is used together with an averaging operator to obtain the macroscopic equation, the mean equations (the Navier–Stokes equation in kinetic theory). In the case of an incompressible ow (the procedure can be easily extended to the compressible case), it is straightforward to prove, using Eqs. (219) and (239), that pE veriAes the following di8erential equation: 1 92 9p E 9p E 9 = − [Ai (x; V; P ) pE ] + [C0 +pE ] ; + Vi 9t 9xi 9Vi 2 9Vi 2
(243)
where Ai (x; V; P ) is given by Ai (x; V; P ) = −
1 9P + Di (x; V) : 9xi
(244)
In order to obtain the (mean) Aeld equations, a standard procedure is used, in analogy with the derivations which can be found in kinetic theory [20,21], as explained at the beginning of the section. Given any function H(V) (it is one component of the tensor of order n; Vi1 Vi2 : : : Vin ), the following expectation : (averaging operator) is deAned H(t; x) = H(U(t; x)) = H(V)pE (t; x; V) d V : (245) This corresponds to a simpliAed case (since is constant) of the deAnition of mean variables already used in Eq. (224). Let us multiply Eq. (243) by H(V) (denoted H for the sake of simplicity) and apply the previously deAned operator. With
9p E 9H E E [HAi p ]; H → 0 ; p (246) Vi →±∞ 9Vi 9Vi (by construction pE and 9pE = 9Vi converge to zero when, at least one component of the velocity goes to inAnity, Vi → ±∞) and if we assume that all generalized integrals converge, after some algebra, one can obtain the following equation:
2 9 9 H 9 9H 1 H + Vi H = Ai + C0 + ; (247) 9t 9xi 9Vi 2 9Vi 2
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
81
provided that uniform convergence is achieved (so that one can write 9= 9t = 9= 9t and 9= 9xi = 9= 9xi ). By choosing H = 1, the mean continuity equation, 9Ui = 9xi = 0 is obtained, and with H = Vi , Eq. (247) yields the mean momentum (Navier–Stokes) equation, 9Ui 1 9P 9Ui 9ui uj + =− : (248) + Uj 9t 9xj 9xj 9xi
The mean Aeld equations of higher order cannot be obtained directly by the procedure which has just been introduced. As a matter of fact, a change of coordinates in phase space is necessary, Vi = Ui (x; t) + vi (velocity is represented in terms of the uctuating component, v = v(V; U ). By noticing that pE (t; x; v) d v = pE (t; x; V) d V, the mean moment of order n is deAned by n ui1 : : : uin (t; x) = vik pfE (t; x; v) d v : (249) k=1
The procedure developed so far with a function H(V) is now repeated with another function H(v) (it is one component the tensor of order n; vi1 vi2 : : : vin ), whose expected value is deAned by H(u(t; x) = H(v)pE (t; x; v) d v : (250) Eq. (243) is Arst re-written is the new phase space. The time and spatial derivatives become (the velocity derivatives are not changed) 9pE (V) 9pE (v) 9Ui 9pE (v) − ; (251) = 9t 9t 9t 9vi 9pE (V) 9pE (v) 9Uj 9pE (v) = − 9xi 9xi 9xi 9vj
(252)
and consequently, the Fokker–Planck veriAed by pE (t; x; v) is 1 92 d pE 9p E 9 = − [Ai (x; V; P )pE ] + [C0 +pE ] + vi dt 9xi 9vi 2 9vi2 +
9Uj 9pE dUi 9pE + vi ; d t 9vi 9xi 9vj
(253)
where d = d t = 9= 9t + Ui 9= 9xi . Let us multiply the previous equation by H(v) and make the same assumptions as in the derivation of Eq. (247) (that is Eq. (246) is veriAed with pE (v), all generalized integrals converge and uniform convergence is achieved). By doing so and after some algebra, one can write
2 1 d 9 9H 9 H H + vi H = Ai + C0 + dt 9xi 9vi 2 9vi2
9Uj 9 dUi 9H − (vi H) : (254) − dt 9vi 9xi 9vj
82
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Inserting H = 1 or vi in the previous equation, one obtains the mean continuity equation and the mean Navier–Stokes equations, respectively. With H = vi vj , the transport equation for the Reynolds stresses becomes 9ui uj 9ui uj uk 9Uj 9ui uj 9Ui + = −ui uk − uj uk + Uk 9t 9xk 9xk 9xk 9xk
+ Gik uj uk + Gjk ui uk + C0 +ij :
(255)
6.7.2. Trajectory point of view Pursuing our trajectory point of view, we present a derivation directly from the particle stochastic di8erential equations. This is a nice exercise for the application of stochastic calculus. In the pdf point of view, the normalization condition of pE yields immediately the zero divergence equation. In the particle point of view, the condition of mass conservation is directly enforced to calculate the mean pressure-gradient. Therefore, by construction, we have 9Ui = 9xi = 0. The mean momentum equation is simply obtained by applying the averaging operator to the particle velocity equation (236b) d Ui = −
1 9P dt : 9xi
(256)
Using the relation between the instantaneous substantial derivative and its Eulerian counterpart, d · = d t = 9 · = 9t + Uj 9 · = 9xj , we obtain 9Ui 1 9P 9Ui 9ui uj + =− : + Uj 9t 9xj 9xj 9xi
(257)
Thus, the high Reynolds form of the mean Navier–Stokes equation is satisAed. This should not be too surprising since convection is treated without approximation by the Lagrangian point of view and since the mean pressure-gradient, which represents the mean value of the acceleration of a uid particle, is properly taken into account in Eq. (236b). Contrary to the Eulerian approach, no model is needed for the Reynolds stress tensor. The second-order equations are more interesting. First of all, one has to write the instantaneous equations for the uctuating velocity components along a particle trajectory. This is done by writing ui = Ui − Ui and consequently d ui d Ui 9Ui 9Ui 9Ui − uj − : (258) = + Uj dt dt 9t 9xj 9xj We now write the equation in an incremental form to properly handle the stochastic terms, which using the mean Navier–Stokes equation is 9ui uk 9Ui d ui = d t − uk d t + Gik uk d t + C0 + d Wi : (259) 9xk 9xk The Arst two terms on the rhs are exact and are independent of the form of the stochastic model. It is seen that, for non-homogeneous turbulence, the mean value of the uctuating velocity increments is not zero. Failing to include that term is equivalent to mishandling the mean pressure-gradient term. This may lead to non-physical e8ects, such as spurious drifts
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
83
unfortunately met in most two-phase ow models, see Section 7.5.2. The di8erent SDEs are deAned in the Itˆo sense and the derivatives of the products ui uj are obtained from Itˆo’s formula d (ui uj ) = ui d uj + uj d ui + C0 + d tij :
(260)
The mean second-order equations are then 9ui uj 9ui uj uk 9Uj 9ui uj 9Ui + = −ui uk − uj uk + Uk 9t 9xk 9xk 9xk 9xk + Gik uj uk + Gjk ui uk + C0 +ij :
(261)
The mean convective terms and the triple correlations arise naturally from the average of the derivative along the instantaneous trajectory and are therefore exact. The turbulence production terms on the rhs stem also from convection and, more precisely, from the second term on the rhs of Eq. (259). Since these terms come from exact manipulation, the production terms are treated without approximation. The e8ect of the model appears on the second line and is seen to correspond to a model for the pressure rate of strain correlation Hij . By using the decomposition of the matrix G already introduced + 1 3 Gij = − (262) + C0 ij + Gija ; 2 4 k a more explicit form is obtained 9ui uj 9ui uj uk 9Uj 9ui uj 9Ui + = −ui uk − uj uk + Uk 9t 9xk 9xk 9xk 9xk 3 + 2 2 a a ui uj − kij − +ij : + Gik uj uk + Gjk ui uk − 1 + C0 (263) 2 k 3 3 Several remarks can be made at that stage. The Arst one is that these second-order equations are the Arst level in the moment hierarchy where the model (through the matrix G) manifests itself. This re ects the fact that the stochastic model represents a model for the joint e8ects of the uctuating pressure gradient and the uctuating viscous term. The Arst mean equation where these e8ects appear is the second-order one through the mean work performed by these forces. The particular form of the second-order equation is of course dependent on the expression of the matrix G or G a . For di8erent forms, we get di8erent models of Hij while with the simplest form, G a = 0, the Rotta model (simple return-to-isotropy term) is retrieved. Through di8erent constitutive laws for G a , one can get various turbulence models. This connection between Lagrangian stochastic equations and Reynolds stress modelling has been comprehensively detailed in some papers [10,61] and these works are referred to for further details. The mean equation picture is completed by writing the equation satisAed by the mean frequency rate !. Following the same derivation, one gets 9! 9uj ! 9! + Uj + = −!2 (S! − h) : (264) 9xj 9xj 9t 6.7.3. Compressible ;ows Mean equations for variable density and compressible ows are derived from the stochastic equations (provided the necessary supplementary models for scalars and other quantities entering
84
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
the chosen models are introduced) with the same procedure. In that case, the proper quantity to handle is the Eulerian mass density function F E which satisAes the same equation as the Lagrangian transitional pdf. The mean values obtained from this mdf are the Favre-averaged means deAned by ˜ (t; x)H(t; x) = H(V; )F E (t; x; V; ) d V d : (265) 6.8. Physical and information contents In this paragraph, we provide a short summary of the main ideas and we discuss physical aspects of the modelling steps as well as the information content of present models. It is important to be aware of the relation between a choice of a state vector and of a stochastic model on the one hand and the corresponding loss of information on the other hand. This point is discussed in the last part of this subsection. 6.8.1. Summary of the modelling steps Brie y speaking, we can merge the Lagrangian and the trajectory point of view and say that the exact instantaneous equations are replaced by modelled ones (stochastic models) which still describe the instantaneous behaviour. This is expressed by the following sketch: + + d xi = Ui d t d xi = Ui d t + ⇒ : (266) 1 9P 1 9P d t + Di d t + C0 + d Wi + QUi dt d Ui+ = − d Ui = − 9xi 9xi The complete model consists therefore in a stochastic di8usion process in terms of the state vector Z = (x; U) (the stand-alone model contains an equation for the dissipation rate and the vector state is extended to Z = (x; U; +) but we leave out this last model in the present section). The trajectories of the process are continuous and a linear law is used for the drift vector Di as a function of the instantaneous particle velocities. This translates, in turbulence modelling, classical arguments from Statistical Physics. However, the matrices G and G a can be any function of mean quantities and the resulting pressure rate of strain model in the second-order equations can be of any form. The precise expression for G a is not provided by the present derivation and a constitutive law still has to be assumed to get a closed expression. The preceding section has shown that there is a strong connection between the present class of Lagrangian stochastic models and second-order equations [7] and this connection is even often put forward as a justiAcation of the Lagrangian models. In the present document, another route has been followed. We have worked our way from the Navier–Stokes equations by directly injecting a model for the instantaneous variables which are considered. This is indeed a mesoscopic model in agreement with the general approach described at the beginning of this section. We have used the Kolmogorov theory to help us make a physically based choice of the variables retained in the state vector and to suggest that, at high Reynolds numbers, the uid particle acceleration was the good variable to eliminate.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
85
6.8.2. Physical content We can try to outline the physics accounted for or left out of the present approach by reformulating the modelling steps with a slightly di8erent language. The uid particle acceleration A is in fact split into three contributions, 1 9P A=− (267) + G(U − U) + : !"# 9x !" # As Al
The sum of the Arst two terms on the rhs, denoted by Al may be regarded as representing the pressure-gradient due to the large scales. In the present one-point framework, this term has to be modelled and is here subdivided into a mean (in the spirit of Reynolds decomposition) and a return-to-the-mean term. This is also a weakness of present one-point formulations since a satisfactory model for the rapid pressure term remains an open issue. A better idea would probably be to try to calculate directly these large scale part of the interactions between particles, in the spirit of the decomposition performed in large eddy simulations (LES). Proposals along that line of reasoning and suggestions on how to couple present methods with other particle methods such as S.P.H. will be put forward in Section 9.3.2. Yet, in spite of their current limitations, present stochastic models have a crucial advantage over classical approaches including the more-advanced LES method: they explicitly simulate all the scales at a given point within the ow and not simply the Arst two moments as Reynolds stress models, see Section 6.3, or just the larger scales as in LES models. This is manifested by the explicit simulation of As = for each particle and not just their indirect in uence on the average terms, such as A or Ai Uj . The direct consequence is that local source terms and particularly those that depend upon the entire range of scales, such as chemical reactive source terms, are treated without approximation, see Section 6.5.3. On the contrary, even LES methods in spite of their interest for large scales predictions are faced with the same di@cult closure requirement for the small scales. Doing a poor job at that stage, or neglecting the sub-grid scalar uctuations on the source terms, can result in predictions that can be completely inaccurate as was shown recently in simple mixing ows [62,63]. In the language of Statistical Physics, the present modelling step, Eq. (266), can be seen as a kind of mean-Aeld approximation. We have replaced the real problem of N interacting particles by N particles that interact with a mean Aeld. The particles do not interact directly but create a mean Aeld or ‘potential’ that is applied to each one. It is also called a ‘weak interaction’ approximation, an expression which highlights already some of its limitations in turbulence modelling. The last term in the acceleration decomposition, , represents small scale uctuations. At high-Reynolds, is treated as a fast process and is replaced by a white-noise term. Indeed, since it represents small scale e8ects, it can be seen as a (real) process whose time scale is of the order of the Kolmogorov time scale $@ and whose variance is from Kolmogorov hypotheses + ()2 ∼ : (268) $@ Therefore, we have when Re → ∞ $@ → 0; ()2 → ∞ with ()2 × $@ → D = + :
(269)
86
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
This is fully consistent with the reasoning presented in Section 4 to justify the shift from a real process to a white-noise process. Such a move is therefore justiAed and mathematically-supported. Furthermore, the Anite trace which is left when the fast process is eliminated is the expression of one keystone of Kolmogorov theory which states that, as → 0 or Re → ∞, the dissipation rate of turbulent kinetic energy tends towards a Anite and non-zero limit. The closure model for is written even in the general case of non-homogeneous turbulence. In that case, the model is local in space dt → C0 +(t; x) d W : (270) Re→∞
This is also in line with the ideas of Section 4 and with the notion of slaving variables. The fast variable is slaved by the slow variable, here the position x, and is replaced by its limit value. This is one way to express the ideas of the (local) equilibrium of the small scales. In summary, we can say that there is about as much physics in present pdf models as in Reynolds-stress models, as far as the predictions of mean velocity and=or temperature Aelds are concerned. The main interest is to provide more local information on the ow and, in particular, on small scales. On the one hand, phenomena governed by small scales as well as large scales are still correctly treated and deviations from Gaussianity can be captured (since the information contains higher-order moments). On the other hand, there is no instantaneous Aeld in these methods and, as such, cannot pretend to simulate coherent structures. These questions will be addressed again in Sections 9.3.1 and 9.3.2. Numerical examples are presented below in Section 6.9 in order to illustrate the previous points by numerical outcomes and to give an idea of how the method works in practice. 6.8.3. Information content It is also useful to emphasize the information content of present models: what is the kind of information available and what is lost? Indeed, a modelling step, such as Eq. (270), in which a di8erentiable process is replaced by a non-di8erentiable process (a white noise) has important consequences. From the considerations developed in Section 4, it is clear that such a move implies an irreversible loss of information. In our case, acceleration is eliminated from the state vector which is valid provided the system (a uid particle) is observed on a ‘large enough time scale’, that is provided we have $@ d t T :
(271)
As a consequence, terms related to the acceleration or governed by what happens in the range d t $@ in which the model is not valid may not be treated correctly. A good example is the behaviour of the autocorrelation near the origin. This quantity is of importance in dispersion models for two-phase ows (it will be discussed in Section 7 and in particular in Section 7.5.3). Let us consider, for the sake of simplicity, homogeneous stationary turbulence when present models have the simple generic form √ U (272) d U = − d t + < d t with < d t = K d W : TL The notations have been chosen so as to agree with those used in Section 4.3. From the results given in that section, we know that when the limit of a white-noise term is taken as in
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Eq. (270), the autocorrelation is an exponential function s U (t0 )U (t0 + s) = exp − : Rf (s) = U 2 (t0 ) T Then, the derivative of Rf (s) at the origin is d Rf (s) 1 =− ; Rf (0) = ds T s=0
87
(273)
(274)
instead of zero as it should be for a real process, see Section 5. However, this is not really a new element and it cannot be used to criticize the Langevin model. Indeed, the slope of Rf (s) when s → 0 is governed by the acceleration, or in other words by the behaviour of the increments of uid velocities over time steps in the range d t $@ , that is outside the range of validity of the model. This can be seen by considering increments of Rf (s) around s = 0 for small values of s 1 1 <(0)U (0) U (0) d U (s) = − d s + ds : d Rf (s) = Rf (s) − Rf (0) = (275) 2 U T U 2 √ When < is treated as a white noise (details on the acceleration are lost), i.e. < d s = K d W , then from the rules of stochastic calculus, we have 1 1 <(0)U (0) = 0 ⇒ d Rf (s) = − d s or Rf (0) = − : (276) T T Yet, when acceleration remains a di8erentiable process, the rules of classical calculus apply. From the conservation of energy, U 2 we have dU 2 = −2
U 2
d t + 2
U 2
(277) T T and, in that case, we obtain the correct derivative for Rf (s), Rf (0) = 0. In consequence, if one is interested in the behaviour of Rf (t) near the origin, then particle acceleration must be properly handled in the modelling steps. From the discussion in Section 4.3, a satisfactory answer is to include the acceleration in the state vector and to reject the model to the time rate of change of the acceleration. For example, in the last process presented in Section 4.3, Eqs. (109), is not a white noise but a coloured noise. Thus, the acceleration exists (U is now a di8erentiable process). Such a model may be useful to simulate uid velocity even when d t $@ by assuming that the characteristic time scale of , that is $ in Eqs. (109), is equal to the Kolmogorov time scale $ = $@ . From the results of Section 4, the form of the velocity autocorrelation is 1 $ (278) e−s=T − e−s=$ ; Rf (s) = 1 − $=T T which does respect the two constraints Rf (0) = 0 and Rf (0) = 0. It is also clear from the above expression, that the correct limit is a question of how the limit is taken when both s and $ go to zero. If for a Axed s, we Arst take the limit $ → 0, then we retrieve the Langevin model for uid velocities and in consequence the slope is not correct. On the other hand, if we take Arst the limit s → 0 for a Axed $ which means that we Arst calculate Rf (0) when uid acceleration still exist, and then $ → 0, we have the correct limit.
88
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
The point emphasized by this example is that one should be aware of the resulting loss of information implied by the necessary modelling steps such as in Eq. (270). A clear physical analysis must then be performed beforehand in order to identify what variables are needed for speciAc issues. 6.9. Numerical examples In the last part of this section, we present and discuss numerical simulations based on pdf models. These numerical examples illustrate how the method works in practice and the kind of information that can be extracted from present models. Present pdf or particle stochastic models can be applied di8erently. First of all, pdf models can be coupled to conventional Eulerian solvers in various ways. In some applications, the pdf approach is limited to scalars variables (in order to treat reaction sources terms correctly) while the Aelds of mean velocity, Reynolds stresses and mean pressure are calculated by solving their PDEs (partial di8erential equations) on a mesh using classical numerical methods such as Anite volume. Thus, an Eulerian solver provides a particle (or pdf) solver with these mean Aelds, and in return the particle solver sends back the value of the mean source terms and of the mean density to the Eulerian solver. In other numerical methods, the moment and pdf approaches are also coupled for the solution of the dynamical variables. For example, one can compute the mean Navier–Stokes equation with a classical Eulerian solver, but where the Reynolds stresses are calculated from the particle set instead of being modelled as in k–+ models. This coupled numerical method was applied by Haworth and El Tahry in an interesting industrial application [64]. A new coupling procedure has recently been put forward [65], in which the Eulerian and Lagrangian solvers are coupled in a consistent way (the introduction of this article provides an up-to-date overview of the various coupling strategies). The coupling of pdf models to Eulerian solvers is not a necessity and a second class of numerical methods makes use only of particles. There is no Eulerian solver in the background and everything (all the necessary or useful Aelds) are computed from the set of particles. This is called a stand-alone pdf code. This method is by construction fully consistent since there is only one approach (the Lagrangian one using stochastic particles). From a strictly practical point of view, this stand-alone approach may not be the most e@cient in terms of computational costs [65]. However, a stand-alone simulation represents a more stringent numerical test compared to coupled approaches. Since there is no Eulerian calculation performed with classical models, any numerical bias or ill-posed numerical algorithm that could be present in the pdf solver is clearly revealed whereas it could remain hidden in a coupled numerical solution. The simulations presented below have been obtained with such a stand-alone approach. Various applications have been described by Pope and di8erent co-workers over the years. These are summarized in a speciAc chapter (Chapter 12) of Pope’s textbook on Turbulent Flows [27]. A detailed analysis of numerical errors and performances of the related stand-alone code has been given by Xu and Pope recently [56]. Another stand-alone code has been developed by the Arst author of this paper in collaboration with Jacek Pozorski (IMP, Polish Academy of Science, Gdansk) and the simulations described below have been obtained with this code. Details on the numerical algorithm have been provided in a recent publication along with one speciAc application [66]. A comprehensive discussion of the overall algorithm or of speciAc numerical
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
89
Fig. 9. Weak approximation of the pdf.
steps is not within the scope of the present paper (which is devoted mainly to two-phase ows), but it may be useful to give the main lines: • the ow is represented by a large number of particles. These particles correspond to samples
of the one-point pdf, or represent di8erent realizations of the ow,
• a number of variables (position, velocity, turbulent frequency, scalars, etc.) are attached to
each particle. The particles are statistical models of real uid particles and the time evolution equations of particle variables are (generally) stochastic equations. • the meaningful quantities are statistical averages and are calculated locally as ensemble averages from the set of particles found in a small volume around the point of interest, see Fig. 9. This corresponds to a weak approximation of the underlying one-point pdf, as indicated in Sections 2.1.2 and 2.10, with the exact pdf being approximated by N
1 p( ) pN ( ) = ( i − ) ; N
(279)
i=1
where i is the value of the variable attached to the particle labelled i and corresponding sample-space variable.
is the
For a given application, the domain is Arst covered with a mesh and is thus divided into small cells. Although the stand-alone approach is in spirit a grid-free method, a mesh is introduced for practical reasons. The mesh is used to locate particles and to calculate statistical averages from the set of particles present in each cell (or perhaps in neighbouring cells also depending upon the chosen method) and also to compute the mean pressure Aeld required to satisfy the mean continuity equation (which is therefore satisAed at the level of the size of a cell) [66]. The global numerical method falls therefore in the category of Monte Carlo particle=mesh approach. A large number of particles are distributed continuously in the domain. At each time step, particle properties are updated by integrating the stochastic di8erential equations in time. Boundary conditions are applied and particles may leave or enter the domain depending upon the case considered. Then, particle locations and velocities are corrected so as to ensure the
90
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 10. Sketch of a mixing layer.
mean continuity equation [66]. Finally, statistical quantities, such as the mean velocity Aeld, Reynolds stresses,: : : ; are computed in each cell as local ensemble averages. 6.9.1. Mixing layer The Arst example described is a mixing layer, see Fig. 10. This is a typical free shear ow which has been well studied over the years. A mixing layer is a turbulent ow obtained by mixing two streams of di8erent velocities, say Ul and Uh (with Ul 6 Uh ). The characteristic scales are indicated in Fig. 10: a longitudinal scale L which scales as x, a transverse scale l, a velocity scale Us = Uh − Ul and a velocity scale for turbulence u. We deAne the characteristic width of the ow by = y0:95 − y0:05 where y0:05 is the location at which we have U (y) = Ul + 0:05Us and y0:95 the location at which we have U (y) = Ul + 0:95Us . As other free shear ows, mixing layers are evolving ows and the transverse distributions of the di8erent variables change with the downstream position. However, an experimentally supported assumption is to regard these ows as self-preserving: the di8erent variables retain the same form and become independent of the longitudinal coordinate x once they are scaled by the corresponding scales (which grow as a function of x). For a mixing layer, the driving force Us remains constant. The turbulent velocity scale u satisAes u=Us = O(=L)1=2 . The characteristic width of the ow grows linearly with x and consequently the ratio =L is constant. Then, u is a constant fraction of Us and is also constant. Experimental evidence indicate that the entrainment parameter U + Uh d S= l (280) Uh − Ul d x is constant S 0:056. We present some results for a mixing layer with Ul = 1 m= s and Uh = 1:5 m= s. Further details can be found in a proceeding paper which describes the computation results [59]. The ratio Ul =Uh = 0:67 is close to one, and thus the present case corresponds approximatively to a temporal mixing layer. For this calculation, a speciAc version of the general algorithm was used. Computations were performed with the boundary-layer algorithm [6] in which particles represent a Axed amount of axial momentum instead of a Axed amount of mass and are marched downstream until a self-similar state is reached. The numbers of particles simulated varied from
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
91
Fig. 11. ProAle of the mean axial velocity. Comparison with the experiments of Wygnanski and Fiedler (•), Patel (?) and Champagne et al. ( ).
about 15 000 at the beginning of the computations to about 100 000 at the end. The overall calculation took around 30 –35 mns for 1000 time steps on a 40 MEGAFLOPS HP workstation. It must be noted that the pdf model used in the calculations correspond to the reAned Langevin model [58] in which the di8usion coe@cient in the velocity equation is written in terms of the instantaneous dissipation rate instead of the mean value. The detailed form of this model is more complicated compared to the models presented in Section 6.6 with many additional variables that increase the computational costs with respect to the simpler SLM (simple Langevin model). The calculated spreading rate d = d x is found to be 0:036 whereas the experimental formula predicts a value of 0:037. It is Arst assessed that a self-similar state is reached by checking that proAles plotted as a function of @ = y= collapse on the same curves. Then, the proAles of mean quantities are compared to the experimental measurements of Patel [67], Champagne et al. [68] and Wygnanski and Fiedler [69]. The mean velocity and the non-zero components of the Reynolds stress tensor are shown in Figs. 11 and 12. Higher-order moments which are not accessible for classical turbulence models are easily calculated from the set of particles with nearly no extra computational costs. For example, triple correlations and skewness and atness factors of the velocity components are displayed in Figs. 13 and 14. The following Agures illustrate the type of information that can be extracted from the model. Fig. 15 is a scatter plot of the instantaneous streamwise velocity of a selection of particles (of the order of 10 000 particles). The low-speed and the high-speed sides are clearly visible; particles are in a laminar state and have nearly the same velocity which collapses on a single line. In the core of the mixing layer, the ow is turbulent and instantaneous velocities are scattered. At a given lateral location, the spread is representative of the axial velocity variance uu1=2 . The next Agure is very instructive (Fig. 16). It displays a scatter plot of the instantaneous value of the logarithm of the relaxation rate log ! for the same group of particles. At the centre of the mixing layer (@ = 0), values of the random variable are spread around a central value. Only one group is visible and the level of log ! is a Arst hint that they are all turbulent particles. At the edges of the layer (for example, @ = y= = 0:15), a di8erent behaviour is evident: there is clearly a bimodal distribution. Particles which are located in the corresponding cell can be divided into two groups. The Arst group is made up of particles with roughly the same level of
92
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 12. ProAles of the Reynolds stress components. Comparison with the experiments of Wygnanski and Fiedler (•), Patel (?) and Champagne et al. ( ).
Fig. 13. Various third-order moments. Comparison with the experiments of Wygnanski and Fiedler (•). The solid line is to be compared with the Alled symbol while the dotted line is to be compared with the empty symbols.
log ! than those which are in the centre of the layer. Compared with the particles which are in
the centre, less particles constitute this Arst subset. The second subset is made up of particles having the same low value of !, ! 0, indicative of laminar particles. These two subsets represent turbulent and laminar particles, respectively, and their simultaneous presence is the
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 14. Skewness and Champagne et al. ( ).
93
atness factors. Comparison with the experiments of Wygnanski and Fiedler (•) and
Fig. 15. Scatter plot of axial instantaneous particle velocities. Fig. 16. Scatter plot of the logarithm of the instantaneous relaxation rate of an ensemble of particles.
translation of the intermittent nature of the ow. In order to bring out this intermittent behaviour in a more quantitative way, the pdf of the local relaxation rate can be derived. Two cells have been considered, one in the centre of the mixing layer where the uid is fully turbulent and one at the edge of the layer where the ow is highly intermittent. By considering the ensemble of particles present in these cells and by sorting the value of ! into bins, it is possible to plot the pdf of the random variable !. Fig. 17 presents the result in the centre of the layer.
94
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 17. Distribution of the pdf of the dissipation rate ! in the centre of the mixing layer. Fig. 18. Distribution of the pdf of the dissipation rate ! at the edge of the mixing layer.
Only one group of particles is present (turbulent particles, < 1) and the distribution follows a log-normal law. Such an outcome should not be regarded as a prediction of the model. Indeed, the model is built upon the reAned Kolmogorov hypotheses [58] and therefore assumes that the relaxation rate is lognormally distributed. However, this Agure re ects an important physical behaviour: at one location, even when the ow is fully turbulent, dissipation is not given by a simple value + but is a random variable. In a homogeneous Aeld, instantaneous snapshots of the dissipation would reveal regions of high turbulent activity (high value of +) concentrated in small volumes of the ow (low probability of these events). This spotty picture becomes more evident as the Reynolds number increases and is the signature of internal intermittency. Another type of behaviour is brought out by drawing similar results with the cell near one edge of the layer (Fig. 18). As described above, there are two groups of particles. Particles belonging to the laminar subset have all the value ! 0 therefore building a Dirac at the origin. Particles in the turbulent subset have scattered values and give rise to the second part of the curve. Although not obvious from the graph, the instantaneous values within the turbulent subset follow a log-normal law. The typical log-normal distribution is here attened since turbulent particles represent a small fraction of the whole set. The relative importance of the two distributions (Dirac versus log-normal) manifests the relative importance of the two subsets and represents directly the intermittent factor <. It is here a result of the present approach and proves that the model is able to simulate external intermittency. The ability of pdf models to simulate naturally external intermittency, and to give access to conditional statistics (statistics calculated only with ‘turbulent’ particles) as well as unconditional statistics is an attractive feature. A detailed physical analysis of the external intermittency coe@cient and its relation with variables of the model (in particular the dissipation-weighted kinetic energy) was proposed in [59]. The comparison between calculated and experimental intermittency coe@cients is shown in Fig. 19. Other results showing the deviation of the pdf of velocity components from Gaussianity were also presented in the same work [59].
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
95
Fig. 19. External intermittency coe@cient. Comparison with the experiments of Wygnanski and Fiedler (•). The solid line represents a direct computation based on the fraction of turbulent particles in each cell. The dotted line is the evaluation of < based on the expression the ratio k= k˜ of the kinetic energy over the dissipation-weighted kinetic energy.
Fig. 20. Sketch of the turbulent channel ow with an indication of typical boundary conditions.
6.9.2. Channel ;ow Another class of ow concern wall-bounded turbulent ows. A typical example is the channel ow. A sketch of the geometry is given in Fig. 20. Compared to free shear- ows, computations of wall-boundary conditions require proper particle boundary conditions. These conditions have been developed in a recently published paper [66]. In this work, a complete 2D algorithm was detailed and results obtained on a high Reynolds channel ow have been presented including an analysis of numerical errors and the importance of variance reduction techniques. The independence of results with respect to the choice of the mesh and to the choice of the number of particles per cell was also demonstrated [66].
96
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 21. Mean streamwise velocity and Reynolds stress as a function of the normalized cross-stream coordinate.
Time-averaged results are presented below. They represent a typical pdf calculation of a high Reynolds channel ow (Re = 120 000 based on the channel half-width and the bulk velocity, or Re$ = 4800 based on the friction velocity) and compared to the experimental results of Comte–Bellot. Numerical computations took 5 mns to compute 1000 time steps with 25 000 particles with a 10 × 25 mesh on a HP785=C3000 workstation. Since detailed numerical results are well described in [66], we limit ourselves to two main outcomes. These results have been obtained with the SLM model instead of the RLM (reAned Langevin model) model used for the mixing layer case and the equations of the model correspond to the models described in Section 6.6. The corresponding second-order model is the Rotta model which includes only a return to isotropy term. This is not the best model to use to obtain the most satisfactory agreement with experimental data. However, the purpose of the computations presented here is more to illustrate the approach and to validate the overall numerical algorithm rather than to obtain the best possible results. The plots in Fig. 21 present mean streamwise velocity as well as components of the Reynolds stress tensor Rij . The limits obtained in the simulations for the values of the Reynolds stress components at the wall are in line with the analytical values derived from the choice of the model in the logarithmic layer [66]. Higher-order statistics have also been computed, for example the skewness and the atness of velocity components which are useful to analyse the deviations of velocity pdfs from Gaussianity. They are shown in Fig. 22. 6.9.3. Heated channel ;ow Finally, we report computational results for scalar variables. The Arst two cases presented above concerned only dynamical variables. The interest in pdf models increases when more complex physics come into play, since their inclusion represent only a small overhead. For
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
97
Fig. 22. Skewness and atness factors of velocity components as a function of the normalized wall distance. Comparison with the experiment of Comte–Bellot.
Fig. 23. Sketch of the heated channel ow with an indication of typical boundary conditions.
example, to simulate a turbulent ow with heat exchange one has to extend the particle state vector to scalar variables for which modelled time-evolution equations have to be written. Typical models for scalars, as well as related issues, are discussed in Section 9.3.3. As a direct follow-up of previous calculations, we present below results for a heated channel ow where temperature is treated as a passive scalar with constant heat ux at the walls. A sketch of the problem with particle boundary conditions is shown in Fig. 23. This calculation is a recent computation obtained with a full velocity-scalar pdf stand-alone approach and with the numerical code developed with Jacek Pozorski (IMP, Polish Academy of Science, Gdansk). Results were Arst presented in a conference poster paper [70]. Yet, the
98
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
Fig. 24. Temperature pdf in the core region of the heated=cooled channel. Left: scatter plot of instantaneous particle temperatures. Right: resulting pdf and histogram.
Fig. 25. Temperature pdfs in the core region of the heated=cooled channel. Comparison of computed pdf (solid line) and Gaussian pdf (dotted line).
case is still being investigated and we only describe here preliminary results. Fig. 24 represents a scatter plot of instantaneous particle temperature values in a small volume located near the core region of the channel. The pdf corresponding to this temperature distributions is deduced simply by bin-counting and is also shown in Fig. 24. It is clear from this result that the pdf is non-Gaussian and this is further revealed in Fig. 25 by comparing directly the calculated pdf with the Gaussian form. It is seen that the model captures a deviation from Gaussian statistics in this non-homogeneous case.
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
99
7. One-point particle pdf models in two-phase 7ows Most of the necessary preliminary work has been performed and we are in a position to develop the pdf approach to two-phase ows. Indeed, the mathematical tools were explained in Section 2, the relations between various state vectors and the hierarchies of pdf equations have been revealed in Section 3 and the physical meaning of using SDEs has been proposed in Section 4. Furthermore, Section 6 has shown how to apply these concepts for single-phase pdf modelling, already introducing all the ideas behind the probabilistic description of turbulent ows. All these tools and results will now be used directly. The pdf approach to two-phase ow modelling will be divided into two sections, the present one and the following one. The following one will deal with the complete formalism for the two phases, the uid and the particle phases, while the present section adopts a pdf point of view only for the particle phase. This is done for two main reasons. First of all, this two-step approach towards the complete pdf formalism greatly simpliAes the theoretical and probabilistic tools. We will then be able to provide an in-depth discussion of some of the physical issues involved without being impeded by the mathematical framework. A second reason is that, by making this choice, the approach presented in this section falls into the rather classical Lagrangian, also called particle-tracking approach for which a vast literature exists, see for example one of the latest reviews by Stock [12] and the references quoted inside. In this literature, the modelling problem is addressed directly through an heuristic use of stochastic models and without making much use of the mathematical background. Although this short-cut method may allow one to obtain quickly numerical answers to certain problems, it is believed that paying not enough attention to the theoretical framework has several drawbacks and quite often leads to severe shortcomings. For instance, a lack of understanding on the assumptions behind stochastic modelling brings about confusion on the nature of the approach. Moreover, the fact that many models actually belong to the same class is missed and these models are believed to be di8erent (for example, the presentation of Markov chains versus random walks as if they were completely di8erent mathematical objects or the fact that current Lagrangian models are pdf models is not understood). Finally and even more importantly, some di@culties are created that actually do not exist (one such example is the notion of the so-called spurious drifts, see Section 7.5.2) while important open questions do not receive enough attention. With respect to that context, the objectives of the present section are (i) to precise what a pdf description is, as well as its relation with macroscopic approaches and its interest, (ii) to give an idea of the physical content of present state-of-the-art models and to emphasize the real modelling issues, (iii) to show that, once mathematical properties and the physical assumptions behind stochastic models are clearly explained, a number of di@culties disappear, (iv) to present numerical examples that illustrate how pdf models work in practice and the kind of information they provide.
100
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
7.1. Fundamental equations and modelling approaches The case of dispersed two-phase ows is now considered, i.e. ows where a continuous phase (a gas or a liquid) is carrying discrete particles (solid particles, droplets, bubbles, etc.), so that, for instance, cases like wavy or slugging two-phase ows are outside the scope of the present work. We will also assume that the characteristic length scale of the discrete particle is much smaller than the characteristic length scale of the continuous phase. New physical points, related to motion of discrete particles in turbulent ows, are introduced in this section and are worth detailing. The form of discrete particle equations is Arst discussed before considering various modelling approaches that can be followed. 7.1.1. Particle equation of motion The starting point of the statistical approach is to write the exact equations of motion of the discrete particles. Loosely speaking, these equations are the equivalent of the fundamental equations for the uid phase, that is the Navier–Stokes equations. Yet, the situation is more complex: the equations of motion for a discrete particle in a turbulent ow Aeld do not have the same level of validity as the Navier–Stokes equations and the exact form remains a subject of current research. This question has a long history. Stokes Arst worked out the resistive drag force for a steady creeping ow in 1851. Basset (1888), Boussinesq (1903) and Oseen (1927) then extended the equation to the case of a particle accelerating in a uid at rest. Tchen, in 1947 [71], was the Arst to generalize this equation for a particle in a turbulent ow. Since then, various contributors have addressed the question and have proposed modiAcations of the expression of the forces, however often in an ad hoc manner. The derivation of the forces was analytically carried out from Arst principles by Maxey and Riley [72] and more rigorously (in our opinion) by Gatignol [73]. In the following, we Arst indicate the main results and then discuss more at length speciAc issues. For negligible relative Reynolds numbers, Rep = dp |Ur |= f based on the particle diameter dp and on the local instantaneous relative velocity between the uid and the particle Ur = Us − Up , the equation has the form d xp = Up ; dt
mp
d Up = F1 + F2 : dt
(281a) (281b)
Here, mp = p ((d3p =6) is the particle mass, F1 accounts for the so-called pressure-gradient and the buoyancy forces whereas F2 stands for the drag, added-mass and Basset forces (d3p DUs (d3p F1 = + (p − f )g ; 6 f Dt 6
(282a)
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
F2 = 3(dp f f (Us − Up ) t (d3p DUs d Up d d$ 3 2 √ − + dp f ( f + : f (Us − Up ) √ 12 Dt dt 2 t−$ −∞ d $
101
(282b)
In these equations, Us = U(xp (t); t) is the uid velocity seen, i.e. the uid velocity sampled along the particle trajectory xp (t), not to be confused with the uid velocity Uf = U(xf (t); t) denoted with the subscript f. The expressions above involve particle and uid accelerations, respectively d Up = d t and DUs =Dt. The precise form of the ‘ uid acceleration’ is still discussed. As will be seen below, that question is not important for heavy particles. It is however of signiAcance for lighter particles, such as bubbles, and will be discussed later on in Section 7.5.6. The division of the forces exerted by the uid on the particle into two terms has some physical meaning. In the course of the mathematical derivation, it is convenient to write the instantaneous velocity Aeld surrounding the particle as the sum of an ‘undisturbed’ Aeld (the velocity Aeld that would exist if the particle was not present) and a ‘disturbance’ Aeld which represents the in uence of the particle. Such a division is possible since dropping the inertia term (thanks to the assumption of small relative Reynolds number) transforms the problem into a linear one. The ‘undisturbed’ Aeld manifests itself through the pressure-gradient and buoyancy forces. On the other hand, the drag, added-mass and Basset forces arise from the ‘disturbance’ Aeld (evidenced by the presence of the relative velocity in the mathematical expressions of the forces) and are expressed as a function of the ‘undisturbed’ velocity at the centre of the particle (which is the precise deAnition of the velocity of the uid seen Us ). It is worth noting that it is the mathematical treatment which Arst yields the above equations which are then physically interpreted. The analytical results stated above rely upon the assumption that Rep 1. For higher Rep , when exact calculation cannot be performed anymore, it is assumed that the di8erent forces already isolated are still present, but with a generalized expression. For example, the drag force is generalized with the help of a drag coe@cient CD . The situation is more complicated for the added-mass and Basset forces. However, one can propose [74] mp
d Up = F1 + F2 ; dt
(283)
where F2 =
1 (d2p CD |Us − Up |(Us − Up ) 2 4 f t (d3p 3 2 √ DUs d Up d d$ + : − + dp f ( f (Us − Up ) √ dt 2 d $ 12 f Dt t−$ −∞
(284)
The added-mass and Basset forces are left unchanged while the drag force has now a quadratic form. The drag coe@cient is an empirical coe@cient that can be estimated through experiments. Various expressions have been put forward, cf. Clift et al. [75], among which an often retained
102
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
form is CD =
24 [1 + 0:15Rep0:687 ]
Rep
0:44
if Rep 6 1000 ;
(285)
if Rep ¿ 1000 :
For heavy particles where p f , the drag and gravity forces are the dominant forces and the particle equation of motion is reduced to d Up 1 (286) = (Us − Up ) + g : dt $p The drag force has been written in this form to bring out the particle relaxation time scale p 4dp $p = : (287) f 3CD |Ur | In Eq. (286), $p appears as the only scale, and is the time necessary for a particle to adjust to uid velocities. In the limit when Rep 1, it is seen from the expression of CD in that range that p d2p $p = (288) f 18 f which is the Stokes value. Even in that limit regime, the particle relaxation time scale is already a non-linear function (here quadratic) of particle properties (the particle diameter dp ). Outside the Stokes regime, the dependence of $p on particle properties and variables, such as Us and Up , is more complicated. Actually, the fact that the added-mass and Basset forces can be neglected with respect to the drag force, when p f , is not so obvious from the expressions of the di8erent forces in Eq. (284) since the particle density is nowhere present on the right-hand side of the equation. This can be justiAed by going back to the analysis of the basic equations, see [73]. The Navier– Stokes equation for the ‘disturbance’ Aeld, say W, around the solid particle is in the creeping regime, 9Wi 1 9p 92 Wi + f 2 : (289) =− 9t f 9xi 9xi The boundary conditions on the particle surface and at inAnity are W = −Uf + Up on the particle surface ; (W; p) → (0; 0)
when r → ∞ :
(290)
The important step is dimensional analysis. We introduce the particle diameter dp as the relevant length scale, Wr as a relevant scale for the relative velocity and $0 as the relevant time scale. With these typical scales, the Navier–Stokes equation is now, using a star to indicate non-dimensional quantities d2p 9p ∗ 9W ∗ 92 W ∗ (Rep St) ∗i = − ∗ + ∗ i ∗ ; Rep St = : (291) 9t 9 xi 9xi 9xi
f $0
J.-P. Minier, E. Peirano / Physics Reports 352 (2001) 1–214
103
Obviously, the choice of the time scale is crucial to the problem. If the time scale $0 is much larger than the viscous time scale, $0 d2p = f , the uid derivative can be neglected and the Navier–Stokes equation is reduced to the Stokes problem around a particle (stationary creeping ow). The force exerted on the particle is then limited to the drag force. On the other hand, if $0 is of the same order of magnitude as the viscous time scale, $0 d2p = f , then the uid derivative must be retained and this leads to the unsteady terms, the added-mass and Basset forces. Consequently, everything relies on the evaluation of the time $0 . Now, we are interested in the particle motion and it seems reasonable to take $0 as the characteristic time for particle velocity changes, that is as the particle relaxation time scale $0 =$p . At this point of the analysis $p is still unknown, but can be worked out as follows. Let us Arst assume that when p f , $p is very large with respect to d2p = f . Then, the added-mass and the Basset forces may be forgotten as explained above. Consequently, the particle equation of motion is reduced to mp or even mp
d Up 1 (d2p = CD |Us − Up |(Us − Up ) + mp g dt 2 4 f
(292)
d Up = 3(dp f f (Us − Up ) + mp g ; dt
(293)
since we are in the small Reynolds number regime. From this equation, as already mentioned, we can isolate the particle relaxation time scale, Eq. (288). Since p f , the relevant time scale $0 that enters the dimensional analysis is indeed much larger than the viscous time scale and the assumption is conArmed. This result is only valid for heavy particles, for when p f , the reasoning fails to single out a particular time scale. The unsteady terms are of the same order of magnitude as the stationary drag force and the only characteristic time is the viscous di8usion time d2p = f . In the following, we mainly limit ourselves to the case of heavy particles (the problem of bubbles or sediments is addressed in Section 7.5.6). The fundamental equations that describe heavy particle motion in turbulent ows are therefore made up by Eqs. (281a), (285), (286) and (287). When particle loading is not too small, other e8ects come into play. Particle–particle interactions or collisions have to be taken into account and the exchange of momentum and of energy between the uid and the particles can be expressed by additional source terms in the Navier–Stokes equations. The above form of the drag coe@cient is essentially for a single particle embedded in a turbulent ow. When the particle loading increases and we move into the realm of dense (but still dispersed) two-phase ows, the drag coe@cient for a particle may be modiAed by the presence and the in uence of other particles. This re ects hydrodynamical forces between discrete particles, see for example [76]. For example, one particle being in the wake of another particle may well experience a somewhat di8erent drag force compared to the upstream one. The hydrodynamical e8ect remains a di@cult