INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDS Int. J. Numer. Meth.

Fluids 2005; 47:11731179 Published online 29 December 2004 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/ d.862

The design and construction of implicit LES models

SUMMARY We o er preliminary thoughts on the design of a modied equation and the construction of a corresponding numerical algorithm, which may be intended for use in an implicit large eddy simulation. The principle of design here is based on ensuring a form for the energy dissipation that is not signicantly dissipative on the resolved scales of the numerical mesh, but is strongly dissipative when the solution is unresolved and so provides strong nonlinear stability in the simulation. The construction process is modelled on the composition (hybridization) of two ux approximations by means of a nonlinear, ow-dependent switch. Published in 2004 by John Wiley & Sons, Ltd.
KEY WORDS:

nonoscillatory methods; implicit turbulence modelling

1. INTRODUCTION Recently there has been an increased interest in implicit large eddy simulation (ILES) where nonoscillatory nite volume (NFV) methods provide e ective modelling of the unresolved dynamics of turbulence in high Reynolds number uid ows [1, 2]. ILES is a simple and computationally e cient approach, and has been successfully applied to complex engineering and geophysical ows. Further, the absence of explicit parameters holds the promise of increased predictiveness in simulations. ILES methods o er a more automatic approach to modelling complex, general systems of equations beyond those where classical turbulence theory is grounded. NFV methods are based on a variety of constructions, both algebraic and geometric. Many of these have been exploited for ILES. The common features of all these schemes are nonlinearity and nite volume approximation. The combination leads to ow-dependent dissipation
Correspondence

E-mail: [email protected] E-mail: [email protected] This article is a U.S. Government

to: L. G. Margolin, MS B218, Los Alamos National Laboratory, Los Alamos, NM, 87545, U.S.A. work and is in the public domain in the U.S.A.

Contract= grant sponsor: U.S. Department of Energy; contract= grant number: W-7405-ENG-36 Contract= grant sponsor: Los Alamos Laboratory Directed Research and Development (LDRD)

Published in 2004 by John Wiley & Sons, Ltd.

Received 27 April 2004 Revised 18 October 2004 Accepted 18 October 2004

1174

L. G. MARGOLIN AND W. J. RIDER

and nonlinear computational stability. Despite this commonality, individual algorithms di er in their details and consequent e ectiveness. In Reference [3] we used modied equation analysis [4] to nd the equivalent partial di erential equations (i.e. modied equations) of several common NFV algorithms, exhibiting the distinct form of the nonlinear dissipative terms in each case. We remark that many (but not all) of these forms can be shown through analysis to correspond to explicit subgrid scale closures used in large eddy simulation (LES). Suppose for the sake of argument that one could devise an optimal method (with its corresponding modied equation) to describe a turbulent ow. Could one construct a numerical algorithm that corresponds to that particular equation? Further, what principles and constraints would one choose to dene optimality? Would these principles and constraints be generic, or be ow specic? This paper is a rst e ort to address these questions. In Section 2, we discuss a particular principle of design, namely that the algorithm must be appropriately dissipative on the resolved length scales of the computation mesh. In Section 3, we describe a general construction process based on a composition of high-order and low-order uxes by a nonlinear, owdependent switch. We examine the result of this reverse engineering process in the context of the one-dimensional Burgers equation for specic choices of the uxes and of the switch. The emphasis in this paper will be on the dissipative character of the e ective model. We provide some discussion in Section 4.

2. ENERGY DISSIPATION The form and magnitude of energy dissipation is a principal concern in the design of a numerical algorithm. Essentially all physical processes, including uid ow, are dissipative. However in many cases of practical interest, the length scales at which energy is dissipated are too small to be resolved in numerical simulation. For example, this is the situation in such diverse problems as large Reynolds number turbulence, and high-speed compressible ows with shocks. This similarity can be exploited in the development of implicit turbulence models. The fact that the dynamics of the dissipative scales are too small to be resolved does not imply that the process of energy dissipation can be ignored. Consider the case of a turbulent uid forced at the large scales of motion. Because of the nonlinearity of the convective processes, energy cascades down in scalee.g. to smaller and smaller eddies. Physically, this cascade process continues until it reaches length scales where molecular viscosity becomes e ective at converting the kinetic energy into heat. However in numerical simulations where the viscous length scales are not resolved, energy will simply build up at the smallest resolved scales if dissipation is absent from the numerical algorithm, producing a qualitatively incorrect solution. This conundrum has been long recognized in the turbulence modelling community. The conventional solution has been to augment the governing equations by explicit terms (i.e. subgrid scale models) whose main function is to model turbulent e ects and in particular to provide appropriate energy dissipation. On the other hand, advocates of the ILES approach rely on the dissipative properties of the numerical algorithm itself to provide the requisite energy dissipation. In ILES, the particular form of the energy dissipation usually results from the imposition of monotonicity constraints. But suppose we wish to take a more
Published in 2004 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2005; 47:11731179

IMPLICIT LES MODELS

1175

active role by designing an algorithm with a particular form of dissipation. On what basis would one attempt to choose that form? The most satisfying basis would be the analytic form of the nite volume-averaged equation. An attempt to derive this form for the one-dimensional Burgers equations was made in Reference [1]. However there remain unresolved issues in the averaging process. Specically, there is a lack of uniqueness in the process that cannot be resolved mathematically, and so requires additional input from physical argument (i.e. thermodynamics). A more practical alternative is to draw on the experience of the turbulence modelling community by prescribing a subgrid scale model as a regularizing term. Many subgrid scale models have been devised and undergone extensive testing and validation against experimental and eld data. Their success and continued use in engineering and geophysical ow simulations imply their utility. At the same time, the existing diversity of such models casts doubt on the existence of an optimal model. To be specic, let us consider the following one-dimensional PDE for velocity u: @2u @u @f(u) + = @t @x @x2 (1)

Here f(u) is a convex ux function and is the physical viscosity. We wish to solve this PDE numerically on a grid with cell size x. We suppose that the viscosity is small in the sense that (U x)= 1 where U is some measure of velocity, e.g. from the initial or boundary conditions of the problem. Then, as described above, (1) is not suitable as a model equation for the numerical simulation and it is necessary to add a term that is dissipative on the scale x in order to ensure numerical stability. We will choose this term from among the explicit subgrid scale models. For the sake of deniteness, here we will focus attention on one subgrid scale model that is both simple and widely used. The Smagorinsky model [5] was rst developed for use in atmospheric simulations. It is closely related to the articial viscosity of von Neumann and Richtmyer [6]. In one dimension the articial stress has the form (C x2 |@u=@x|@u=@x) where C is a dimensionless constant. Then the modied equation that we seek is @ @u @f(u) + = C x2 @t @x @x @u @u @x @x (2)

The specics of the numerical ux f(u) will determine the bounds on the value of C . Here we have omitted the physical viscosity as being inconsequential to the numerical simulation. Multiplying (2) through by the velocity, we derive the associated equation for kinetic energy 2 K=1 2 u @F (u) @K @u + = C x2 @t @x @x
3

+ C x2

@ @u @u u @x @x @x

(3)

where @F=@u u@f=@u. Integrating this equation over the entire mesh clearly shows the dissipative nature of the modied equation.
Published in 2004 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2005; 47:11731179

1176

L. G. MARGOLIN AND W. J. RIDER

3. REVERSE ENGINEERING In this section we describe a process to construct an algorithm for simulating (1) whose modied equation matches a specied form, e.g. Equation (2). We term this process reverse engineering (RE). RE does not necessarily have a unique result. Although we have not required that the RE algorithm be nonoscillatory, it appears reasonable to attempt to model the RE process on similar ideas used to build NFV methods. One such idea is the melding of low-order and high-order uxes (e.g. FCT [7], MPDATA [8], TVD [9]). This melding or hybridization is accomplished by means of a nonlinear, ow-dependent switch . We write the algorithm in the conservation (nite volume) form
n+1 n = uj uj

t n n (f f j 1= 2 ) x j+1=2

(4)

where in common notation the superscript n indicates the time level, the subscript j is a spatial index identifying the centre of the cell, and t is the computational time step. Further, we write the uxes generically as
LO + (1 j+1=2 )fHO f j +1= 2 = j +1= 2 f

(5)

The high-order ux, fHO , is chosen to given accurate solutions in smooth regions of the ow. A common choice is the LaxWendro ux 1 n 1 LW n f j +f j +1 ) j +1= 2 = (f 2 2 @f n n (u u j ); @u j+1

@f t @u x

(6)

where is the dimensionless Courant number. The low-order ux fLO is chosen to ensure su cient energy dissipation in regions of rapidly varying ow. The most common choice is simple upwinding 1 n 1 @f up n n n (uj f j +f j +1 ) +1 uj ) j +1= 2 = (f 2 2 @u (7)

As regards the switch, we will require [0; 1]. This ensures that the composite scheme (4) is stable for at least the more restrictive of the time step conditions associated with the individual ux models F HO and F LO . Furthermore, this limitation of also ensures ux consistency f(uj ; uj+1 ) = f(u) if uj = uj+1 = u

assuming that the individual uxes have this property. We also want 0 as the ow becomes smooth, to ensure an accurate solution in these regions. A simple way to characterize smoothness on the mesh scale is through the dimensionless ratio ( x=u@u=@x). For example, one might dene j+1=2 min 1; C uj+1 uj uj+1 + uj + (8)

where the small (properly dimensional) constant x= t 1015 avoids division by zero. Another e ective form for the switch is ( x(@x=@u)@ 2 u=@x2 ) which closely corresponds to the minmod limiter [3].
Published in 2004 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2005; 47:11731179

IMPLICIT LES MODELS

1177

Let us apply these choices: (6) for the high-order ux, (7) for the low-order ux, and (8) for the switch, to the case of Burgers equation where the ux function f(u) = u2 . In the limit t 0, we nd the modied equation: u2 ) @ @u @( 1 + 2 = x2 @t @x @x
u

1 6

@2u 1 @x2 6

@u @x

C @u @u 4 @x @x

(9)

The terms not present in our original modied equation are related to linear dispersion (which can be removed with a third-order base scheme) and a nonlinear term associated with control volume di erencing. The corresponding equation for kinetic energy is @( 1 u3 ) x2 @ @K + 3 = @t @x 6 @x u2 @2u @u u @x2 @x
2

x2

C @u 4 @x

1 6

@u @x

(10)

which is dissipative for C 2 3. Comparing (10) with (3), we see that they have qualitatively the same form of energy dissipation. Our RE equation does not treat rarefying and compressing regions symmetrically, a modication that is now preferred for the articial viscosity in shocks as well. Also, the RE equation has extra dispersive terms that move energy around, but do not change the dissipative properties of the equation. The e ect of this dispersion has been discussed in Reference [11]. We note that the modied equation (9) is very similar to that for MPDATA [1], even though the motivation for its derivation is much di erent. MPDATA in its basic form [12], as well as our RE method, is sign-preserving but not monotonicity preserving. In our case, this is a direct result of the form for the switch , which is not gauge invariant (i.e. depends explicitly on where the zero of velocity is set). Other forms for the switch that are monotonicity preserving are proposed in Reference [13] through the use of modied equation analysis. MPDATA is derived as a multipass method, where truncation errors are compensated with upwind approximations. In this approach, it is possible to eliminate unwanted dispersive errors, as appear in (9). The elimination of unwanted second-order dispersion in MPDATA is discussed in Reference [14]. Our limiter (8) was designed to emphasize the use of the accurate ux fHO for smooth ow. An alternate idea is to completely turn o any contribution of fLO wherever possible, while strongly dissipating ows that are not su ciently resolved. This implies a more nonlinear switch. One common switch that has this characteristic is based on the monotone limiting such as that used in MUSCL or FCT methods, e.g. a minmod limiter. To illustrate, rst dene the monotone bound f j +1 f j; f j f j 1 ) j +1= 2 = 2 minmod(f (11)

This is a rst-order accurate approximation to the ux. Now we dene a nal ux similarly,
HO f j + minmod( f j) j +1= 2 = f j +1= 2 ; f j +1= 2 f HO where the f j +1= 2 can be any second-order or higher ux.

(12)

For

articial viscosity this is the result of work by Marshall Rosenbluth while at Los Alamos in 1955 [10]. Int. J. Numer. Meth. Fluids 2005; 47:11731179

Published in 2004 by John Wiley & Sons, Ltd.

1178

L. G. MARGOLIN AND W. J. RIDER

4. DISCUSSION In this paper we have explored the design of model equations for ILES, and the construction of a corresponding algorithm. The underlying principle for the design of a model equation is that of ensuring energy dissipation on the smallest resolved scales of the computational mesh. We note that a similar strategy is employed in the variational multiscale methods [15] through the use of a spectral viscosity. In spectral space, the quantication of scale size in the numerical solution is explicit and so it is possible to design a wavenumber dependent viscosity that damps the smallest resolved scales without a ecting the larger scales of motion. The situation is more complex in nite di erence algorithms, where all computational cells have the same length scale, and there is no readily accessible local quantication of scale size in the numerical solution. Instead, in conventional turbulence simulations explicit scaledependent dissipative terms are added to the equations, while in ILES simulations dissipation results from the constraint of preserving monotonicity. In Reference [3] we demonstrated the equivalence of these approaches. In the current paper, we have exploited this equivalence to design new algorithms for modelling turbulent ow by using explicit subgrid scale models to regularize the evolution equations. We note that these subgrid scale models are independent of the physical viscosity of the uid. This is an indication that the details of the small scales of uid motion where physical viscosity is operating do not matter, or more precisely that the small scales are enslaved by the larger scales of motion. It is this enslavement that underlies the success of both LES techniques and of ILES.
ACKNOWLEDGEMENTS

We gratefully acknowledge discussions with Fernando Grinstein and Christer Fureby. Los Alamos National Laboratory is operated by the U.S. Department of Energy under contract W-7405-ENG-36. This work has been supported in part by the Los Alamos Laboratory Directed Research and Development (LDRD) program.
REFERENCES 1. Margolin LG, Rider WJ. A rationale for implicit turbulence modeling. International Journal for Numerical Methods in Fluids 2002; 39:821 841. 2. Margolin LG, Smolarkiewicz PK, Wyszogrodzki AA. Implicit turbulence modeling for high Reynolds number ows. Journal of Fluids Engineering 2002; 124:862 867. 3. Rider WJ, Margolin LG. From numerical analysis to implicit subgrid turbulence modeling. AIAA Paper 20034101, Proceedings of the 16th AIAA CFD Conference, 2326 June 2003, Orlando, FL. 4. Hirt C. Heuristic stability theory for nite di erence equations. Journal of Computational Physics 1968; 2: 339 355. 5. Smagorinsky J. General circulation experiments with primitive equations: I. The basic experiment. Monthly Weather Review 1963; 101:99 164. 6. von Neumann J, Richtmyer RD. A method for the numerical calculation of hydrodynamic shocks. Journal of Applied Physics 1950; 21:232237. 7. Boris JP, Book DL. Flux-corrected transport: 1. SHASTA, a uid transport algorithm that works. Journal of Computational Physics 1973; 11:38 69. 8. Smolarkiewicz PK. A simple positive denite advection scheme with small implicit di usion. Monthly Weather Review 1983; 111:479 486. 9. Sweby PK. High resolution schemes using ux limiters for hyperbolic conservation laws. SIAM Journal on Numerical Analysis 1984; 210:995 1010. 10. Richtmyer RD, Morton KW. Di erence Methods for Initial Value Problems. Wiley: New York, 1967. Published in 2004 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2005; 47:11731179

IMPLICIT LES MODELS

1179

11. LeFloch PG, Natalini R. Conservation laws with vanishing nonlinear di usion and dispersion. Nonlinear Analysis-Theory Methods and Applications 1999; 36:213230. 12. Smolarkiewicz P, Margolin LG. MPDATA: a nite-di erence solver for geophysical ows. Journal of Computational Physics 1998; 140:459 480. 13. Rider WJ, Margolin LG. Simple modications of monotonicity limiters. Journal of Computational Physics 2001; 174:473 488. 14. Margolin LG, Smolarkiewicz PK. Antidi usive velocities for multipass donor cell advection. SIAM Journal on Scientic Computing 1998; 20:907 929. 15. Hughes TJR, Mazzei L, Oberai AA, Wray A. The multiscale formulation of large Eddy simulation: decay of homogeneous isotropic turbulence. Physics of Fluids 2001; 13(2):505 512.

Published in 2004 by John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Fluids 2005; 47:11731179