Nonideal Rayleigh-Taylor Mixing

H. Lim , J. Iwerks , J. Glimm , and D. H. Sharp

Department of Applied Mathematics and Statistics, Stony Brook University, Stony Brook, NY 11794-3600, USA, Computational Science Center, Brookhaven National Laboratory, Upton, NY 11793-6000, USA, and Los Alamos National Laboratory, Los Alamos, NM 87545

Submitted to Proceedings of the National Academy of Sciences of the United States of America

Rayleigh-Taylor mixing is a classical hydrodynamic instability, which occurs when a light uid pushes against a heavy uid. The two main sources of nonideal behavior in Rayleigh-Taylor (RT) mixing are regularizations (physical and numerical) which produce deviations from a pure Euler equation, scale invariant formulation, and nonideal (i.e. experimental) initial conditions. The Kolmogorov theory of turbulence predicts stirring at all length scales for the Euler uid equations without regularization. We interpret mathematical theories of existence and non-uniqueness in this context, and we provide numerical evidence for dependence of the RT mixing rate on nonideal regularizations, in other words indeterminacy when modeled by Euler equations. Operationally, indeterminacy shows up as non unique solutions for RT mixing, parametrized by Schmidt and Prandtl numbers, in the large Reynolds number (Euler equation) limit. Verication and validation evidence is presented for the large eddy simulation algorithm used here. Mesh convergence depends on breaking the nonuniqueness with explicit use of the laminar Schmidt and Prandtl numbers and their turbulent counterparts, dened in terms of subgrid scale models. The dependence of the mixing rate on the Schmidt and Prandtl numbers and other physical parameters will be illustrated. We demonstrate numerically the inuence of initial conditions on the mixing rate. Both the dominant short wavelength initial conditions and long wavelength perturbations are observed to play a role. By examination of two classes of experiments, we observe the absence of a single universal explanation, with long and short wavelength initial conditions, and the various physical and numerical regularizations contributing in different proportions in these two different contexts.

Introduction Rayleigh-Taylor (RT) instability is a classical hydrodynamic instability [4, 34] which occurs when a light uid pushes against a heavy uid. The density contrast is measured dimensionlessly by the Atwood number A = (2 1 )/(2 + 1 ), with i the density of uid i. The mixing rate can be characterized by the dimensionless parameter , dened through the equation h = Agt2 , [1]

initial conditions, inuence , with a relative magnitude that is context dependent. A comparison of many codes [9] for the computation of did not achieve agreement. There was agreement neither with experiment (by a factor of 2), nor internally among the various computations. This discrepancy is attributed by some to long wavelength noise (not measured) in the experiments. The discrepancy among the simulations has increased [3] since the time of this study. The simulation discrepancy is unrelated to the long wavelength noise, which was not included in these simulations. The authors and coworkers [14, 13, 23] observed the importance of minimizing numerical mass diffusion in Rayleigh-Taylor simulations. The above cited papers achieved values for consistent with experiment using a special code (FronTier), based on front tracking, and designed to eliminate numerical mass diffusion and to model correctly physical mass diffusion. Particle methods [16], also known to have low levels of numerical mass diffusion, achieve values of consistent with experiment. A combined experimental-simulation study [27, 28] addressed both the initial condition issue and the numerical mass diffusion, and achieved agreement with experiment. It did not address the relative importance of these two effects. The Kolmogorov theory of turbulence [17] suggests that stirring occurs for the Euler equations in a self similar manner at all length scales. Applying this theory to unregularized simulations of thermal or concentration discontinuities, we expect the discontinuities to be present on all length scales, in a self similar manner. Within a numerical simulation, we thus expect a chaotic mixture of discontinuities, limited only by grid effects. This picture implies a numerical interface I with surface area |I| having a constant value in grid units per numerical grid cell [20], or globally, |I| = O(1)(x)1 . [2]

Such a divergent interface behavior was observed [20] in a mesh renement study of unregularized 2D Richtmyer-Meshkov (RM) unstable (shock accelerated) mixing. The identity (2), combined with common error estimates, yields an L1 error estimate error = Const.|I|x = Const. [3]

where h is the penetration distance of the light uid into the heavy and g is the acceleration. The mixing rapidly becomes turbulent, with Reynolds numbers Re 50, 000 observed in typical experiments. For this reason, simulations based on compressible codes, with hyperbolic CFL restrictions, are unable to resolve the viscous length scales, and such codes are run in an under resolved manner. Such simulations are called large eddy simulations (LES). LES require subgrid scale (SGS) models to describe the effect of the small (subgrid) scales on the large (grid resolved) ones. Models for the Reynolds stress based on eddy viscosity are the most familiar of this type. The main result of this paper is a verication and validation study of a new LES algorithm [21] and an analysis of the dependence of on the Reynolds, Schmidt and Prandtl numbers, and on the initial conditions, in two similar but different and independent experiments. According to the numerical evidence presented here, depends on the Schmidt and Prandtl numbers at high Reynolds numbers, approximating the high Reynolds number limit. Stated more directly, the Euler equation description of RT mixing is indeterminate. For one of the experiments, also depends signicantly on the initial conditions. We conclude that both nonideal effects, regularization and
www.pnas.org

In other words, the solution is not convergent in the L1 norm under mesh renement. We summarize the above analysis by stating that Rayleigh-Taylor mixing, when described by the Euler equations, is indeterminate. However, simple addition of the physical parameters for uid transport is very far from a solution to this problem. They are conventionally set to zero for good reason. For most feasible grids, they are so small as Reserved for Publication Footnotes

PNAS

Issue Date

Volume

Issue Number

17

Growth rate

to have the appearance of little effect on the numerical solution. The resolution to this problem is through the use of subgrid models. Subgrid models aim to determine all effects of the unresolved (subgrid) scales on the resolved ones. In this study, we use dynamic subgrid models [26], which have the advantage of not possessing adjustable parameters. As subgrid models are recognized to be an essential ingredient in the solution, they will no doubt be subject to renewed scrutiny in this turbulent mixing context. Mesh convergence (verication) and agreement with experiment (validation) will be important gures of merit. Any adjustable parameters used for this purpose detract from the success, and leave open the proper setting of problem dependent parameters in regimes for which there are no experiments. The SGS models are also not sufcient; in some cases, modeling of experimental (nonideal) initial conditions is also important. In proposing subgrid scale models, we believe we are combining the best of two numerical traditions, that of turbulence modeling (with SGS models) and sharply resolved numerical gradients (also called capturing methods). We enhance this combination with the use of front tracking to achieve still sharper (physically correct) gradients for the concentrations.

RT RM KH

Table 1. Regime Dependence of Instabilities RM and KH micro-characterization is conjectural Initialization Evolution Macro Observ. Micro Observ. Active Active Sensitive Sensitive Active Passive Insensitive Sensitive Passive Passive Insensitive Sensitive

0.08

0.06

0.04 Experiment 0.02

X X

Mathematical Existence and Non Uniqueness of Solutions The Euler equations are known to have non-unique solutions, starting with the 1993 paper of Schefer [33], and rened in [35, 18]. These mathematically rigorous constructions are not very intuitive, but numerical examples [24, 10], built around bifurcation points for unstable shear layers, are easier to understand. These numerical examples have a different character from the numerical non uniqueness we observe here. Since verication of numerical solutions requires showing that they agree with the correct mathematical solutions of the same equations, we note a problem with verication in the ILES context whereby the mathematical solutions are not unique if the physical regularization is not specied. ILES or MILES (Implicit Large Eddy Simulation) [37] is a branch of numerical code design in which the inuence of subgrid scale regularization (viscosity, mass diffusion, surface tension) is modeled implicitly, as a consequence of cutoff procedures specic to each code. These methods generally aim for accuracy at scales as close to the grid scale as possible, and the accuracy in each of the equations proceeds independently of the other equations, without concern for the relative accuracy of these different cutoffs. Non uniqueness implies that the Euler equation is indeterminate; additional information is needed to specify a unique solution. While the mathematical constructions of non uniqueness offer little insight on how to do this, we proceed from physical and numerical intuition and posit that the limiting process remembers the Schmidt and Prandtl numbers as all transport coefcients converge to zero. In other words, we propose that RT mixing, even when modeled by the Euler equations (with zero transport coefcients) requires specication of a Schmidt and Prandtl number to achieve a unique solution. The numerical evidence presented here supports this picture. A similar non uniqueness for a 1D conservation law with a simple jump initial condition (Riemann problem data) was identied in 1970 [6], but the potential for a broader signicance of this phenomena was overlooked. A similar phenomena occurs for reactive ow, for which the ame speed depends on the Prandtl number, even when modeled by the Euler equation (for which the Prandtl number is indeterminate) [7]. In view of the importance of non uniqueness as a possible obstacle to verication, a deeper understanding of these constructions and proofs would be desirable. A very practical, but open, question is to specify which class of ows are likely to be subject to non uniqueness, and which observables of such ows are subject to non uniqueness. We offer, as a conjecture, the summary in Table 1. For each of three classical instabilities, we classify initial thermal and concentration discontinuities as describing either active or passive scalars, in their
2 www.pnas.org

Simulation (FT) Simulation (TVD) Simulation (alpha group) Cabot & Cook 0.01

0.005

Dimensionless surface tension

Fig. 1. Plot of RT growth rate vs. dimensionless surface tension. Shown are all experimental values not using surfactants, several front tracking simulations [13], with varying levels of surface tension including one without use of surface tension, and several untracked simulations [19, 9, 3] without use of surface tension. We note (a) the excellent agreement of the tracked simulations with experiment, (b) the signicant dependence of the tracked simulation s on surface tension and (c) the discrepancy between the untracked simulations with experiment, with tracked simulations and with each other.

inuence on the turbulence evolution, and we classify observables as sensitive or insensitive according to whether they depend signicantly on the Schmidt and Prandtl numbers (numerical and physical, laminar and turbulent). The sensitive observables are expected to be indeterminate when modeled at the level of the Euler equations (without regularization and without specication of the Schmidt and Prandtl numbers). Macro observables describe the overall scope of the mixing zone (such as the RT ), while micro variables refer for example to the joint temperature and concentration probability distribution functions. Table 1 is based on our numerical evidence and the following intuition: Density contrasts affect the initial RT instability and drive its continued evolution. They affect the initiation of RM instabilities, with the deposition of vorticity on the contact discontinuity at the time of shock passage (or reshock), but between shock passages, the instability basically coasts, with dissipation caused by radiation of energy in sound waves in addition to the effect of viscous forces. In this sense, the concentrations and associated density differences are active scalars for RT (initiation and evolution) and RM (initiation only). As discussed below, the evidence for the RM micro characterization (sensitive vs. insensitive) is ambiguous. The Kelvin Helmholtz (KH) instability has smaller sensitivity to density differences, in that the instability retains a nite strength even when the densities are equal. So we describe the concentration and density as a passive scalar in this case. The intuitive picture which supports the sensitivity of the microvariables is the following thought experiment: Imagine a very high Schmidt number ow (low mass diffusion), with stirring at all length scales. The result (if mixing red and blue uid) would be a ne scale mixture of read and blue with little purple. Now imagine the same stirring, but with a small Schmidt number (high diffusivity).
Footline Author

The result will be mixed species everywhere, that is shades of purple. To connect this picture to the macro variables, we need to examine the instability mechanism. Only with RT does the instability mechanism depend on a density in a manner that persists in time. For KH, the instability persists in the absence of a density difference, so the macro variables are insensitive to density differences. For RM, the instability is sensitive to density differences, but only at an initial time instant, before mass diffusion has taken place. Existence theory, for classical weak solutions, has progressed through various special solutions, but even the function space for general solution existence has not been identied. In [20], we conjectured that the space of measure valued solutions was the natural space for a general existence theory. Existence in this space has recently been proved [11], and so our contribution is to suggest that this is the place to stop, without pursuit of additional regularity in the form of a classical weak solution. Even with this suggestion, the existence theory in not complete, in that the proof does not address convergence of the nonlinear terms. In other words, it is not determined whether the limit of the approximate solutions (the limit solution) actually solves the original equation, or some new, modied equation dened through the limiting process. Still, it is safe to say that a more promising route to mathematical existence has been identied. For the regularized equations, such as the Navier-Stokes equations with full transport terms (for viscosity, heat conductivity and mass diffusion, if multi-species), we expect classical smooth solutions and uniqueness.

Table 2. Validated RT Mixing Simulations. sim References SGS Comments Compressible Simulations, Ideal Initial Conditions 0.0650.07 0.069 [1, 23] No Air-helium 0.050.077 0.066 [32, 36, 13] No Immis., no surfact. 0.058 0.057 [36] Yes Fresh-salt water; Present study Compressible Simulations, Experimental Initial Conditions 0.070 0.011 0.070 [27] Yes Hot-cold water; Present study 0.085 0.005 0.075 [27] Yes Fresh-salt water; Present study Particle Based Simulation, Ideal Initial Conditions 0.51 0.06 [8, 16] No Avg Immis Exp Incompressible Simulations, Experimental Initial Conditions 0.07 0.07 [28] DNS Hot-cold water; 0.07 0.06 [30, 31] No Hot-cold water;

exp

4 3.5

X X X X X X X
Experiment #112 = 0.058 3.75X0.625X15 t_0 = 1.5 = 0.056 3.75X0.625X15 (2X refined mesh) t_0 = 2.5 = 0.057 7.5X1.25X15 t_0 = 4.1 = 0.058 15X2.5X15 t_0 = 3.3 = 0.057

Bubble penetration (cm)

3 2.5 2 1.5 1

Verication and Validation ILES emphasize the quality of the cutoff or resolution separately in each of the equations, each with its own primitive variable, but ignores the ratios of the cutoffs as applied to the different equations in the system. In other words, the balance in the relative accuracy of the momentum equation vs. the energy and concentration equations is not part of ILES. This balance is important in the study of mixing. When not set by physics (the Schmidt and Prandtl numbers), numerical ratios (the numerical Schmidt and Prandtl numbers) govern the solution. Even when physical parameters are used, if the simulation is underresolved, the physical parameters fail to regularize the solution, so the numerical parameters dominate, and the numerical Schmidt number is still important. Not only are such ratios conventionally ignored in an ILES simulation, but they would be difcult to document, as the strength of the cutoff process which replaces the subgrid effects are not characterized by a coefcient that can be interpreted in terms of viscosity, mass diffusion or heat conduction. We approach the numerical solution of the indeterminate Euler equations (or the physically regularized, but numerically underspecied equations) with a physically motivated sequence of approximate solutions. This preferred limit removes non uniqueness in a manner which depends on the specic choice of physical regularization it is modeling. It thus allows verication (mathematical correctness of the numerical solution, often tested as convergence under mesh renement) and validation (agreement with experiment). Verication was addressed in an earlier study of the related Richtmyer-Meshkov (RM) instability, initiated in a shock wave induced uid instability of a density discontinuity layer. In a 2D mesh renement study, we analyzed separately the macroscopic variables, such as the edges of the mixing zone, and the microscopic variables, such as the probability distributions for the temperature and concentration or a chemical reaction rate. The macroscopic variables were found to be mesh convergent, even in the absence of regularization, and not sensitively dependent on regularization physical parameters or algorithmic differences [38, 25]. The microscopic variables were sensitive to physical regularization parameters, but with these specied, also mesh convergent [21]. In the absence of physical regularization, these same microscopic variables were found to be sensitive to algoFootline Author

0.5 0 30

40

50

60

70

Scaled acceleration distance Agt^2 (cm)

Fig. 2. Validation comparison of simulation to experiment, showing the experimental data points and several simulations based on the methods proposed here: subgrid scale models to include subgrid scale effects in an LES, and front tracking to eliminate numerical mass diffusion and to allow physical mass diffusion. The several simulation plots reect enlargement of the statistical model (1/4, 1/2 and full size relative to the experimental data) and renement of the computational mesh (for the 1/4 size domain only). The renement comparisons serve as a verication test as well. The quoted experimental value = 0.058 differs slightly from the value = 0.062 quoted in [36], and was recomputed by the present method from the experimental data for consistency of comparison to our simulations.

rithmic differences, in that different numerical codes gave apparently converged but distinct solutions [25, 20]. We summarize previous RT validation studies before presenting the main validation results for the new methods proposed here. We found excellent agreement between simulations and a miscible experiment [1] and with several immiscible experiments [32, 36]. See Fig. 1 and Table 2. In our simulations, we include the relevant physical transport mechanisms and other scale symmetry breaking physical effects that regularize the simulations. A new verication and validation result, Fig. 2, tests the SGS models proposed here, combined with front tracking, in a 3D RT context. To the authors knowledge, this is the rst successful verication and validation study of RT mixing, using a compressible hydro code and (as LES is a necessity) subgrid scale models to compensate for the unresolved scales. It is the rst such study even for moderate Schmidt numbers Sc = O(1), but here is considered with Sc = 560;
PNAS Issue Date Volume Issue Number 3

it is also the rst high Sc validated RT simulation. It is also, to the authors knowledge, the rst RT simulation to agree with all the experimental data points, and not just their late time aspmptotics. Thus it is a signicant step beyond existing simulations. Fig. 2 displays the experimental data points and several simulation studies. The various simulations explore convergence of the statistical ensemble dened by the initial conditions (ensembles reduced by 1/4 and 1/2 in each transverse direction are compared to the simulation of the full device). Mesh renement by a factor of 2 is presented for the quarter size ensemble. None of these changes makes an appreciable difference in the answer. The initial conditions and physical parameters were chosen to follow the experiments, but (for reasons of computational efciency) we did not try to duplicate the very small amplitude scale of the initial perturbation. Accordingly, we allow an offset in time (chosen by least squares) in the comparison to experimental data. After choosing this offset, the value of is determined by a least squares t to (1), with an experimentally determined offset in the initial value of h. The offset in the initial h is related to the time (or penetration distance) delay for the onset of the self similar regime. Applying this method to the experimental data yields an = 0.058, slightly modied from the value = 0.062 quoted in [36]. No long wavelength components were included in the random initial conditions. We next address validation in the context of the water channel experiments [27, 28]. For these experiments, the initial conditions derive from experimental measurement, and there is no role for th in setting them. With a very small Atwood number and earths gravity g0 as the acceleration force, these experiments have a slow time scale and are strongly incompressible. To achieve a feasible simulation problem, we articially modify the time scale, to introduce still small, but increased, levels of compressibility. We do this by articially modifying the gravity to g = 100g0 , and setting A = 1/3 in the problem. To avoid modifying the physical problem (other than by this increase in compressibility), we make a compensating modication in the viscosity. In this way, the Grashof number, Gr = 2Ag3 / 2 , expressing the ratio of accelerating to viscous forces, is not modied. Here is the kinematic viscosity and = exp is the wavelength of maximum amplitude, as measured experimentally. Similarly, the diffusivity has to be rescaled. In view of the modied viscosity, we modify the diffusivity by keeping the Schmidt number constant. These changes do not affect pure length scales, so that the initial mass diffusion length scale is unchanged and the perturbation wavelengths in the initial conditions are unchanged. th is also unaffected. The initial conditions are taken from experimental measurement, and th is introduced here for interpretation only. The change of time scale produces an overall change in the velocity perturbation amplitudes. Even with this change, the compressible simulation is expensive, and we also reduce the grid resolution below that used by [28]. Thus our simulation is LES; SGS terms are employed. We obtain excellent agreement with experiment. Our results are conrmed by comparison to the earlier validated DNS [28] for this experiment. We are not limited to DNS, and we also simulate the fresh-salt water experiment [27]. This simulation, and the results presented above modeling other fresh-salt water experiments [36] are the rst validated simulation of a high Schmidt number Rayleigh-Taylor experiment. The results are slightly outside of the experimental error bars, but still very good. See Table 2. The use of LES rather than DNS is an advantage in its own right, for reasons of computational efciency. The use of a compressible code is also an advantage. Beyond questions of scientic principle, such as a possible loss of or a basis for universality for the value of , a purpose of simulations of this type is to validate a code. Many problems are compressible and require a compressible code. For this purpose, we also require validation of a compressible code; the incompressible validation, while of interest in its own right; does not serve this purpose.

Indeterminacy and Variation of Regularizing Parameters We observe signicant dependence of on the parameters which govern the regularization, or deviation from the pure Euler equations. The results, summarized in Table 3, conrm the idea that an effort to dene the solution of the Euler equations as a limit of the (presumed unique) solution of the regularized equations is indeterminate. The regularizing parameters are viscosity, mass diffusion (Schmidt number) and an initial mass diffusion length IMD . From these quantities, we compute the wavelength th of the most rapidly growing mode. The dominant initial perturbation length scale is centered about the wavelength th . We use the quarter sized ensemble from Fig. 2, with a resolution of 8 cells per initial wavelength and we vary the regularizing parameters one at a time. Since there is no experimental data for most of the simulations in Table 3, we determine the t offset as a least squares t to an h offset linear h vs. Agt2 relationship. Simulations #1-3 record variation of Re. #2 and #4-5 record variation of the initial mass diffusion layer. The clearest evidence for indeterminacy comes from Sc comparion simulations #2 and #6-8. Fig. 1 also provides evidence for indetermincy.

Initial Conditions vs. Fluid Transport The relative role of initial conditions vs. transport in their inuence on is a major open problem for RT mixing. The dominant wavelength in the rocket rig experiments [32, 36] is well predicted by perturbation theory (dispersion relations), both for miscible and immiscible experiments, and for a variety of different experiments. Dispersion relation length scales are in reasonable agreement with the dominant short wavelengths measured for the water channel experiments also. The short wavelength perturbations may be described by the small amplitude RT theory, for both classes of experiments. The long wavelength perturbations in the Mueschke et al. experiments [27] are at the level of 75% of the primary short wavelength perturbations in amplitude. Initial data for the rocket rig experiments were not recorded, so that we are limited to information that can be extracted from the experimental plates. The immiscible experiments provide clearer data, and we will discuss this case. Starting with the second or third plate, the RT growth has reached a visible amplitude. The data suggests a magnitude of at most 10% for long wavelength perturbations, with two or three exceptions. Experiment #63 [2] had a long wavelength perturbation imposed by a vibrating device and for experiment #104 [36], a long wavelength perturbation was noted experimentally (but not deliberately imposed). Experiment #56 [2] shows a possible long wavelength perturbation. The long wavelength perturbation amplitude for these cases has a magnitude of 15% to 40% in the second or third plate of the perturbed experiments, relative to the dominant short wavelength perturbations. The long wavelength noise appears to be signicantly higher in the splitter plate experiments than in the rocket rig experiments, while the short wavelength perturbations may have a common origin as small amplitude RT modes.

Table 3. Variation of Parameters for RT Mixing.

#1 #2 #3 #4 #5 #6 #7 #8 0.060 0.057 0.062 0.053 0.055 0.057 0.055 0.049

th
0.17 cm 0.21 cm 0.26 cm 0.29 cm 0.15 cm 0.21 cm 0.23 cm 0.27 cm

Sc 560 560 560 560 560 50 7 1

Re 105 5 104 2.5 104 5 104 5 104 5 104 5 104 5 104

IMD
0.35 cm 0.35 cm 0.35 cm 0.7 cm 0.175 cm 0.35 cm 0.35 cm 0.35 cm

www.pnas.org

Footline Author

Table 4. s with and without long wavelength perturbations Perturbed Exp/Sim. Unperturbed Perturbed Unpert. Experimental s #63 [2] #56 [2] 0.069 0.058 #104 [36] #105,114 [36] 0.068 0.072, 0.060 Simulation s [28]; Pres. Study Pres. Study 0.070 0.043

The effect on , compared to unperturbed but otherwise identical or nearly identical experiments was 0.011, 0.006, -0.004 in three comparisons. As a further test of the importance of long wavelength perturbations, we repeated the hot/cold water channel simulations, but with all long wavelength perturbations in the initial conditions eliminated. See Table 4. The effect on was 0.25 ( 35%). These experiments do not support the view that long wavelength perturbations in the data are sufcient to explain the factor of 2 discrepancy between many simulations and virtually all experiments. The dimensionless parameters /th and /IMD that govern the short wavelength perturbations appear to be also important, (See Tables 3, 4.) and must be controlled in RT validation. To better understand the consequences of the various parameter modications being compared, we constructed a sequence of ve simulations, starting with the salt-fresh water experiments [36] (simulation #1) and ending with the water channel experiments [28] (simulation #5). Connecting these is a change of uid parameters, decreasing Sc from 560 to 7 and decreasing Gr from 2.1E6 to 6.9E5 (simulation #2), [28]. This is followed by modication of the parameters governing the ideal initial condtions (simulation #3) and with experimental short wavelength initial conditions but no long wavelength initial perturbations (simulation #4). See Table 5. The experimental evidence suggests that the inuence of long wavelength perturbations on is below what is needed to explain the factor of 2 discrepancy between experiment and many simulations. There is numerical evidence that numerical mass diffusion is able to explain much of this discrepancy. Tracked (without numerical mass diffusion) simulations were compared to untracked ones (with numerical mass diffusion) [12]. The comparison used a time dependent Atwood number A(t) dened within the simulation. The time dependence of A is strictly due to numerical mass diffusion, in that the simulations in question lacked physical mass diffusion. The A(t) of the untracked simulation was roughly a factor of 2 smaller than that for the untracked simulation. Using A(t), a renormalized growth coefcient ren was dened through the equation h(t) = 2ren Z tZ
0 s 0

Additional studies favoring the importance of numerical issues related to transport rather than the long wavelength perturbations in the initial data are the theoretical bubble merger models based on bubble competition [15, 5, 29] (in that they predict experimental values for while assuming ideal initial conditions). Other studies [31, 37] favor the importance of the initial conditions. In [37], a weak perturbation of 592 = 3481 long wavelength modes is shown to result in a factor of 2 increase in . Because the regularizing physics sets the length scale of the short wavelength perturbations, it also in effect sets the strength of the regularization, which has to be judged relative to the length scale on which it operates. For example, if the length scale in the Grashoff number Gr is dened by dispersion relations using only viscosity and acceleration, we obtain a pure number, Gr 2(4)3 3969. In this sense, the universality of the regularizing strength (on its self dened length scale), rather than some universal property of scale invariant physics, is probably at the origin of the somewhat common values, across many different experiments and choice of experimental uids, observed for .

A(s )gds ds .

[4]

Using (4) to dene ren , the tracked and untracked simulations had a common value of ren , but values of differing by a factor of 2. From this fact, we conclude that the numerical mass diffusion, as recorded in the time dependent Atwood number A(t), was sufcient to cause a factor of 2 difference in , and thus explain the largest part of the factor of 2 discrepancy between experiments and many simulations.

Sc Gr /th /IMD
Footline Author

Table 5. Simulations connecting [36] and [28] #1 #2 #3 #4 #5 0.057 0.052 0.043 0.043 0.070 -0.005 -0.009 0.0 +0.028 560 7 7 7 7

2.1E6
1 0.6

6.9E5
1 1.0

6.9E5
2.62 12

6.9E5
2.62 12

6.9E5
2.62 12

RT Validation for ILES Based on the results presented here and in [21, 22], we propose a computational strategy for an ILES code to achieve a preferred solution that is (a) veried (mesh convergent) and (b) validated (in agreement with experiment). Since the presumption in the meaning of ILES is that no SGS model will be added and the numerical mass diffusion is whatever the numerical algorithm gives, it is necessary to increase the numerical viscosity so that the numerical Schmidt number is equal to the physical Schmidt number. Since these codes normally contain a parameter for physical viscosity, this parameter can be increased to achieve the desired effect, and the needed increment in viscosity can be called numerical, articial, or turbulent viscosity. Such a terminology is consistent with accepted usage, as the numerical mass diffusion and the required numerical modication to the physical viscosity both scale with x. The resulting viscous length scale, which is now some multiple of x, replaces the mesh length as a measure of the accuracy of the simulation. In other words, by losing some fraction or multiple of a decade of agreement with Kolmogorov theory in the k5/3 turbulent energy spectrum, agreement with laboratory experiments for can be achieved. In keeping with our prior results [21, 22], we expect this computational strategy to yield convergence for the probability distributions for temperature, concentration and reaction rates as well. As the mesh is further rened, the lost decades of Kolmogorov agreement are recovered, while agreement with experiment is preserved. This is the meaning of validation, and since it proceeds with convergence under mesh renement, one test of verication will be achieved. Since the limit process contains supplementary information not present in the Euler equations (the physical Schmidt number to guide the setting of the numerical Schmidt number), the process is consistent with the mathematically established non uniqueness theory. A practical difculty in the approach outlined above is the absence of accepted calibration of the implicitly dened ILES cutoffs in terms of viscosity, mass diffusion and heat conduction. If they are truly to serve such a purpose, such a calibration would be required. In this sense, an ILES Reynolds number, for example, lacks an agreed upon denition. The dynamic SGS models used here are functions of the nonlinear terms in the Navier-Stokes equations, which are the same as the nonlinear terms in the Euler equations. In other words, the SGS have no direct dependence on the physical transport terms, and depend on them only indirectly in that they are dened by the computed solution which has these terms in its governing equations. It could be the case that the introduction of turbulent Schmidt and Prandtl numbers, as free or adjustable parameters, will be helpful for the SGS LES models. In this case the models lose their dynamic quality, acquiring adjustable parameters.
PNAS Issue Date Volume Issue Number 5

Conclusions Regularization of simulations for Rayleigh-Taylor mixing is achieved by the addition of transport terms to governing equations. From a practical numerical perspective, with the necessity of LES, we add SGS terms to the equations. This approach will not end controversies between alternate methods, but we hope it will advance the ideas, so that the remaining discussions focus productively, for example on effective SGS models, just as they now focus on effective CFD algorithms. Accurate modeling of initial conditions may also be needed; the importance of this step will vary from one experiment to another. In this spirit, we have proposed a preferred sequence of approximate solutions which meets standards of verication, validation, physical modeling and mathematics based on fundamental principles. We propose how such a preferred limit could be achieved in the context of conventional ILES codes, without use of front tracking. We comment on the science and mathematics of RT mixing, including the relative role of transport (physical and numerical) vs. long wavelength perturbations in setting experimentally observed values of and we note the role of parameters /th and /IMD governing the short wavelength perturbations. For RM mixing, the numerical evidence is not conclusive on the Sc dependence in the large Re limit. But for nite Re, with nite values of transport, RM, in its microscopic variables such as reaction
1. A. Banerjee and M. J. Andrews. Statistically steady measurements of RayleighTaylor mixing in a gas channel. Physics of uids (1994), 18(3), 2006. (c)2006 The Thomson Corporation, IEE. 2. K. D. Burrows, V. S. Smeeton, and D. L. Youngs. Experimental investigation of turbulent mixing by Rayleigh-Taylor instability, II. AWE Report Number 0 22/84, 1984. 3. William Cabot and Andrew Cook. Reynolds number effects on Rayleigh-Taylor instability with possible implications for type Ia supernovae. Nature Physics, 2:562568, 2006. 4. S. Chandrasekhar. Hydrodynamic and Hydromagnetic Stability. Oxford University Press, Oxford, 1961. 5. B. Cheng, J. Glimm, and D. H. Sharp. A 3-D RNG bubble merger model for RayleighTaylor mixing. Chaos, 12:267274, 2002. 6. C. Conley and J. Smoller. Viscosity matrices for two-dimensional nonlinear hyperbolic systems. Comm. Pure Appl. Math., 23:867884, 1970. 7. R. Courant and K. Friedrichs. Supersonic Flow and Shock Waves. Springer-Verlag, New York, 1967. 8. G. Dimonte and M. Schneider. Density ratio dependence of Rayleigh-Taylor mixing for sustained and impulsive acceleration histories. Phys. Fluids, 12:304321, 2000. 9. G. Dimonte, D. L. Youngs, A. Dimits, S. Weber, M. Marinak, S. Wunsch, C. Garsi, A. Robinson, M. Andrews, P. Ramaprabhu, A. C. Calder, B. Fryxell, J. Bielle, L. Dursi, P. MacNiece, K. Olson, P. Ricker, R. Rosner, F. Timmes, H. Tubo, Y.-N. Young, and M. Zingale. A comparative study of the turbulent Rayleigh-Taylor instability using high-resolution three-dimensional numerical simulations: The alpha-group collaboration. Phys. Fluids, 16:16681693, 2004. 10. Volker Elling. A possible counterexample to wellposedness of entopy solutions and to godunov convergence. Math. Comp., 76:17211733, 2006. 11. W. Gangbo and M. Westerdickenberg. Optimal transport for the sustem of isentropic euler equations. Comm. PDEs, 2009. In Press. 12. E. George and J. Glimm. Self similarity of Rayleigh-Taylor mixing rates. Phys. Fluids, 17:113, 2005. Stony Brook University Preprint number SUNYSB-AMS-04-05. 13. E. George, J. Glimm, X. L. Li, Y. H. Li, and X. F. Liu. The inuence of scale-breaking phenomena on turbulent mixing rates. Phys. Rev. E, 73:016304, 2006. 14. E. George, J. Glimm, X. L. Li, A. Marchese, and Z. L. Xu. A comparison of experimental, theoretical, and numerical simulation Rayleigh-Taylor mixing rates. Proc. National Academy of Sci., 99:25872592, 2002. 15. J. Glimm and D. H. Sharp. Chaotic mixing as a renormalization group xed point. Phys. Rev. Lett., 64:21372139, 1990. 16. K. Kadau, C. Rosenblatt, J. L. Barber, T. C. Germann, Z. Huang, P. Carles, B. L. Holian, and B. J. Alder. The importance of uctuations in uid mixing. Proc. Nat. Acad. Sci. U.S.A., 104:77417746, 2007. 17. A. N. Kolmogorov. C. R. Acad. Sci., 13:301538, 1941. 18. Camillo De Lellis and Laszelo Szekelyhidi. On admissibility criteria for weak solutions of the Euler equations. Archive for Rational Mechanics and Applications, 2009. To appear. See also arXiv:0712.3288v. 19. X.-L. Li. Study of three dimensional Rayleigh-Taylor instability in compressible uids through level set method and parallel computation. Phys. Fluids A, 5:19041913, 1993. 20. H. Lim, Y. Yu, J. Glimm, X.-L. Li, and D. H. Sharp. Chaos, transport, and mesh convergence for uid mixing. Acta Mathematicae Applicatae Sinica, 24:355368, 2008. Stony Brook University Preprint SUNYSB-AMS-07-09 Los Alamos National Laboratory preprint number LA-UR-08-0068.

rate, does show sensitivity to the transport coefcient ratios (Sc and the Prandtl number P r) [22]. Our evidence suggests that for the Euler equations, the RT is indeterminate. It depends on regularizing physics and on nonideal initial conditions. Perhaps the most important conclusion of this study is that we have achieved, for the rst time, RT mixing simulations in agreement with experimental, theoretical and mathematical facts, using methods applicable to problems of practical interest. This simulation capability is based on an improved scientic understanding.
ACKNOWLEDGMENTS. This work was supported in part by the U.S. Department of Energy, including grants DE-FC02-06-ER25779 and DE-FG52-06NA26205. This material is based upon work supported by the Department of Energy [National Nuclear Security Administration] under Award Number NA28614. The simulations reported here were performed in part on the Galaxy linux cluster in the Department of Applied Mathematics and Statistics, Stony Brook University, and in part on New York Blue, the BG/L computer operated jointly by Stony Brook University and BNL. This manuscript has been co-authored by Brookhaven Science Associates, LLC, under Contract No. DE-AC02-98CH1-886 with the U.S. Department of energy. The United States Government retains, and the publisher, by accepting this article for publication, acknowledges, a world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for the United States Government purposes.

21. H. Lim, Y. Yu, J. Glimm, X. L. Li, and D. H. Sharp. Subgrid models for mass and thermal diffusion in turbulent mixing. Phys. Fluids, 2009. Stony Brook Preprint SUNYSB-AMS-08-07 and Los Alamos National Laboratory Preprint LA-UR 08-07725; Submitted for Publication. 22. H. Lim, Y. Yu, J. Glimm, and D. H. Sharp. Nearly discontinuous chaotic mixing. J. High Energy Physics, 2008. Stony Brook University Preprint SUNYSB-AMS-09-02 and Los Alamos National Laboratory preprint number LA-UR-09-01364; Submitted for Publication. 23. X. F. Liu, E. George, W. Bo, and J. Glimm. Turbulent mixing with physical mass diffusion. Phys. Rev. E, 73:056301, 2006. 24. M. Lopez, H. Lopez, J. Lowengrub, and Y. Zheng. Numerical evidence of nonuniqueness in the evolution of vortex sheets. ESAIM:Mathematical Modeling and Numerical Analysis, 40:225237, 2006. 25. T. O. Masser. Breaking Temperature Equilibrium in Mixed Cell Hydrodynamics. Ph.d. thesis, State University of New York at Stony Brook, 2007. 26. P. Moin, K. Squires, W. Cabot, and S. Lee. A dynamic subgrid-scale model for compressible turbulence and scalar transport. Phys. Fluids, A3:27462757, 1991. 27. Nicholas Mueschke. Experimental and numerical study of molecular mixing dynamics in Rayleigh-Taylor unstable ows. Ph.d. thesis, Texas A and M University, 2008. 28. Nicholas Mueschke and Oleg Schilling. Investigation of Rayleigh-Taylor turbulence and mixing using direct numerical simulation with experimentally measured initial conditions. i. Comparison to experimental data. Physics of Fluids, 21:014106 119, 2009. 29. D. Oron, L. Arazi, D. Kartoon, A. Rikanati, U. Alon, and D. Shvarts. Dimensionality dependence of the Rayleigh-Taylor and Richtmyer-Meshkov instability late-time scaling laws. Phys. of Plasmas, 8:28832890, 2001. 30. P. Ramaprabhu and M. Andrews. Experimental investigation of Rayleigh-Taylor mixing at small atwood numbers. J. Fluid Mech., 502:233271, 2004. 31. P. Ramaprabhu and M. Andrews. On the initialization of Rayleigh-Taylor simulations. Phys. Fluids, 16, 2004. 32. K. I. Read. Experimental investigation of turbulent mixing by Rayleigh-Taylor instability. Physica D, 12:4558, 1984. 33. V. Scheffer. An inviscid ow with compact support in space-time. J. Geom. Anal., 3:343401, 1993. 34. D. H. Sharp. An overview of Rayleigh-Taylor instability. Physica D, 12:318, 1984. 35. A. Shnirelman. On the nonuniqueness of weak solutions of the Euler equations. Comm. Pure Appl. Math., 50:12611286, 1997. 36. V. S. Smeeton and D. L. Youngs. Experimental investigation of turbulent mixing by Rayleigh-Taylor instability (part 3). AWE Report Number 0 35/87, 1987. 37. D. L. Youngs. Application of MILES to Rayleigh-Taylor and Richtmyer-Meshkov mixing. Technical report, 2003. Technical Report 4102, 16th AIAA Computational Fluid Dynamics Conference. 38. Y. Yu, M. Zhao, T. Lee, N. Pestieau, W. Bo, J. Glimm, and J. W. Grove. Uncertainty quantication for chaotic computational uid dynamics. J. Comput. Phys., 217:200 216, 2006. Stony Brook Preprint number SB-AMS-05-16 and LANL preprint number LA-UR-05-6212.

www.pnas.org

Footline Author