Chaos, Transport, and Mesh Convergence for Fluid Mixing

H. Lim,1 Y. Yu,1 J. Glimm,1, 2 X.-L. Li,1 and D. H. Sharp3

1 Department

of Applied Mathematics and Statistics,

Stony Brook University, Stony Brook, NY 11794-3600, USA

2 Computational

Science Center, Brookhaven National Laboratory, Upton, NY 11793-6000, USA

3 Los

Alamos National Laboratory, Los Alamos, NM (Dated: January 30, 2008)

Chaotic mixing of distinct uids produces a convoluted structure to the interface separating these uids. For miscible uids (as considered here), this interface is dened as a 50% mass concentration isosurface. For shock wave induced (RichtmyerMeshkov) instabilities, we nd the interface to be increasingly complex as the computational mesh is rened. This interfacial chaos is cut o by viscosity, or by the computational mesh if the Kolmogorov scale is small relative to the mesh. In a regime of converged interface statistics, we then examine mixing, i.e. concentration statistics, regularized by mass diusion. For Schmidt numbers signicantly larger than unity, typical of a liquid or dense plasma, additional mesh renement is normally needed to overcome numerical mass diusion and to achieve a converged solution of the mixing problem. However, with the benet of front tracking and with an algorithm that allows limited interface diusion, we can assure convergence uniformly in the Schmidt number. We show that dierent solutions result from variation of the Schmidt number. We propose subgrid viscosity and mass diusion parameterizations which might allow converged solutions at realistic grid levels.
PACS numbers: 47.27.wj, 47.27.tb, 47.27.-i

Keywords: Schmidt number, Mass Diusion, Turbulence, Multiphase ow

2
I. INTRODUCTION A. Overview

Acceleration driven turbulent mixing is a classical hydrodynamical instability, in which acceleration is directed across a uid interface separating distinct uids of dierent densities23 . We are concerned here with impulsive acceleration produced by a shock wave passing through the uids. This problem is known as the Richtmyer-Meshkov (RM) instability. We consider a circular geometry, with a converging circular shock at the outer edge, and inside this, two uids separated by a perturbed circular interface. The problem was previously described in detail12,14,26 . We consider the problem in two dimensions, in order to pursue conveniently issues of mesh renement. Initial and late time simulation density plots are shown in Fig. 1. The uid interface, at late time, is volume lling. The transport coecients (viscosity, mass diusion, and heat conductivity) are given dimensionlessly as the Reynolds number Re = U L/, the Schmidt number Sc = /, and the Prandtl number P r = /. Here is the kinematic viscosity, the mass diusivity and the thermal diusion rate. Throughout this paper, thermal diusion is set to zero. U and L are characteristic velocity and length scales. Dilute gases have Schmidt numbers of the order of unity or smaller, while liquids and dense plasmas typically have larger Schmidt numbers. We show that the solution depends on the Schmidt number. With under resolved physical transport mechanisms, their numerical analogues (numerical mass diusion, heat conduction and numerical or articial viscosity) play the role of the missing or under resolved physical variables, and select the solution. Computer codes then apparently converge to one of the nonunique solutions. The ability of the microphysics at the smallest turbulent scales (and those below this scale, limited by a diusive length scale) to aect the macroscopic ow occurs with turbulent combustion18 , where details of atomic level mixing aect stoichiometry and temperature, thus the local ame speeds and the macroscopic ow. From the point of

FIG. 1: Initial (left) and late time (right) density plot for the Richtmyer-Meshkov uid instability under study in this paper.

view of turbulent combustion, a converged solution should yield a converged probability density function (pdf), as a function of space and time, for the joint distribution of the species concentrations and the temperature. The sensitive dependence of the temperature and species concentration pdfs is interface related and is amplied by the chaotic nature of the volume lling interface.

4
B. The Flow Instability Problem

In the problem considered here, see Fig. 1, the ow is dominated by a single strong shock wave, starting at the outer edge of the computational domain (a half circle). The shock passes through the interface separating the two uids, proceeds to the origin, reects there and expands outward, recrossing the interface region and nally exiting at the outer boundary. The interface region, due to the shock induced instability, expands into a mixing zone, which has a very complex structure. Especially after the second passage of the shock (the reshock or reected shock passage), the mixing zone becomes highly chaotic. The inner and outer edges of the mixing zone are dened in terms of 5% and 95% volume fraction contours, after a spatial average over the circular symmetry variable. The mixing zone is then dened as the region between these inner and outer edges. The software which captures the space time trajectory of these waves in the numerical solution is known as a wave lter7,8,26 . A space time plot of the shock trajectories and mixing zone edges is shown in Fig. 2. The numerical solutions are by the front tracking FronTier algorithm6 . This algorithm uses a Godunov nite dierence solver based on the MUSCL algorithm4,24 , and a sharp (tracked) interface to eliminate, or optionally13 , to limit mass diusion across the interface. Thus the FronTier numerical Schmidt number is , and it allows simulation of any desired (physical) Schmidt or Prandtl number. From the point of view of turbulence modeling, the simulations can be described as under resolved or resolved direct numerical simulation (DNS). Specically, no subgrid algorithms are used to represent the unresolved scales. In order to explore grid and convergence eects on simulated turbulence and mixing in high Schmidt number ows, we articially specify non-physical (high) values for viscosity. In this way we explore the range of convergence to DNS for a single problem. We have already observed11 that the interface for the problem under study is chaotic, with length proportional to x1 , with respect to its mesh (non) convergence (i.e. rate of divergence) properties. This fact is demonstrated in Fig. 3, where the length / area ratio

Time

0 0

20

40

60

80

100

120

10

15

20

Radius

FIG. 2: Space time (r, t) contours of the primary waves, as detected by the wave lter algorithm. These are the inward (direct) and outward (reected) shock waves and the inner and outer edges of the mixing zone, all detected within a single rotational averaging window, in this case [45o , 0o ].

is shown in physical units (left frame) and mesh units (right frame); transport coecients have been set to zero. From Fig. 3, we observe that somewhat after reshock, the interface length, in mesh units, occupies a constant ratio to the mesh area of the mixing zone. We call this ratio the mesh level surface fraction. Its value is approximately time independent (about 30%), after a transient period following the second shock passage. We note here and in many later plots, some lost of mesh level complexity in the nest grid simulations. Further mesh renement studies will be needed to determine the evolution of mesh level

20

0.6

interface length/mixing zone volume

15

grid length/grid area

200 x 400 400 x 800 800 x 1600 1600 x 3200

200 x 400 400 x 800 800 x 1600 1600 x 3200

0.4

10

0.2

20

40

60

80

100

120

20

40

60

80

100

120

time

time

FIG. 3: Plot of the interface length divided by the mixing zone area vs. time. Left: Length and volume measured in physical units. Right: Length and volume measured in mesh units ([physical length / physical area] x). Results for four mesh levels are displayed.

complexity under continued mesh renement. In any case, at the grid levels attained here, the interface is mesh volume lling, cuto by the mesh, and highly complex or chaotic in nature.

C.

Organization of Paper

The goal of this paper is to study mixing at the atomic scale, and its dependence on transport parameters (viscosity and diusivity). According to conventional ideas, viscosity15 (after a time delay2 for the initiation of turbulence) limits the vorticity at high wave numbers. At high Schmidt numbers, and at still higher wave numbers, the scaling exponent for the concentration uctuations is predicted to change from an ObukovCorrsin exponent to a lower Batchelor exponent1,22 , before being suppressed by mass diusion. These ideas form the basis of the stretched vortex subgrid model10,21 . Presumably, there will be a time delay for the applicability1 of the Batchelor spectrum, which is based on an assumed statistically steady state theory for the concentration uctuations. We follow a dierent approach to this problem, based on DNS to determine the con-

7 centration pdfs for converged solutions and to parameterize diusive subgrid models for large eddy simulations (LES). In this sense, our program can be related to conventional ideas in the modeling of turbulence18 . In Sec. II, we explore the role of viscosity to limit the velocity uctuations and small scale eddies. The Atwood number of the ow is dened as A = (2 1 )/(2 + 1 ). Initially A = 0.72 while at time of reshock, A 0.26. This non negligible value for A indicates that density uctuations couple into the uid equations, and thus that the concentration is not accurately regarded as a passive scalar. We report evidence of mesh convergence of the statistics describing interfacial complexity and the ne scale features of the mix, for simulations in a DNS regime. Evidence indicating a sensitivity of temperature to details of physical and numerical modeling for the ow problem considered here was reported earlier12,14 . Sec. III studies mass diusion to determine the concentration pdfs. A nal Sec. IV contains some speculations regarding mathematical existence theories and comments regarding subgrid models to allow converged calculations on realistic meshes.

II.

TURBULENCE, CUTOFFS AND MESH CONVERGENCE

We set a (time dependent) length scale L to be the width of the mixing zone, and the velocity scale U to be the turbulent velocity U = The Kolmogorov scale is K = ( 3 / )1/4 . Here = where Sij is the strain rate tensor, Sij = 1 2 ui uj + xj xi . (2) 2 Sij v 2 . The angle brackets denote an angular (or in principle ensemble) average. We also dene Remesh = U x/. is the dissipation rate,
2 2

(1)

We plot Re and the mesh Kolmogorov scale Kmesh = K /x vs. Remesh for the simulations analyzed here, in Fig 4. The condition Kmesh 1 is generally taken as the denition of a DNS.

8
10
6

101
200x400 400x800 800x1600 1600x3200 200x400 400x800 800x1600 1600x3200

105
4

100

10

Kmesh

Re

10

10

-1

10

10
10
1

-2

10 -1 10

10

10

10

10

10

10-3 -1 10

10

10

Remesh

Remesh

10

10

10

FIG. 4: Left: Re vs. Remesh at t = 90 for several mesh levels. Right: Kmesh vs. Remesh . Both plots are virtually independent of Sc for the range of Sc considered. Some of the simulations at each grid level are resolved below the Kolmogorov scale, and are resolved DNS. Others, on the route to resolution, are under resolved DNS. A. Convergence of Macroscopic Flow Observables

The objective here is to compute the large scale solution features accurately. This includes the trajectories of the principal waves, as illustrated in Fig. 2, and other macro solution features, such as the mean densities and velocities of each phase within each homogeneous solution region between the principal waves. This objective is related to the systematic convergence study carried out earlier26 . In that study, we found statistical convergence for many mean ow variables, to dene what we call the macroscopic description of the ow. In Fig. 5, we plot the relative wave error dened in terms of the mixing zone edge positions, for a variety of mesh levels and transport coecients. The error (or discrepancy) is determined by comparison of the simulation to a ne grid (3200 1600) simulation having zero transport coecients. The reported discrepancy is thus a mixture of mesh errors (except for the nest grid) and discrepancies associated with modication of the transport coecients from a nominal value (zero). We consider convergence in this macro sense as a function of Remesh for each value of x. From Fig. 5, we see that for the three grid levels shown, the smallest converged value

9
0.2
400x800 Sc=0.1 400x800 Sc=1 400x800 Sc=10 800x1600 Sc=0.1 800x1600 Sc=1 800x1600 Sc=10 1600x3200 Sc=0.1 1600x3200 Sc=1 1600x3200 Sc=10

relative wave error

0.15

0.1

0.05

0 10-1

100

101

Remesh

102

103

104

FIG. 5: Time integrated relative errors in the mixing zone edge locations as compared to a ne grid, zero transport simulation.

of Remesh progresses from about 30 to 2 as the mesh is rened. Some of the simulations shown (coarse grids, to the far left of Fig. 5) are clearly not convergent and are eliminated from subsequent analysis. With a macro converged simulation having Remesh = 2, we turn to Fig. 4 and observe that on the nest grid, the corresponding Kmesh = 2, indicating that the simulation is clearly DNS. To examine in more detail the eects of modied transport on the accuracy of the simulation, as compared to an ideal case with zero transport, we consider in Table I a range of additional variables. We consider mean density and radial velocity data. The mean angular velocity is essentially zero. Energy data is not presented, as it was observed12,14 that temperature is sensitive to details of numerical and physical modeling; we expect a similar dependence for energy. Energy and temperature will be considered in a separate study in which the Prandtl number plays a role. We consider Sc = 10 but not Sc = 1.0 or Sc = 0.1. The density and velocity distributions depend on the Schmidt number, and the high Schmidt number cases are the most comparable to the ideal case (computed with ideal parameters, i.e. with zero numerical mass diusion). Since any physical problem must have a nite Schmidt number

10 and other transport coecients, the discrepancy between the ideal (zero transport) and the nite Schmidt simulations probably reects ideal simulation errors compared to an (infeasible) DNS simulation with molecular values for the transport coecients. We cannot fully determine which of the two solutions is better. We only present data showing that the dierence is small for Sc = 10. We consider separately the errors in the mixed regions and the single phase regions, singly or doubly shocked. Each simulation thus denes ve regions in space time, to be analyzed separately, with boundaries as illustrated in Fig. 2. Comparing two simulations, we consider the space time points which lie in a common region for both simulations. Call these points R, and call the uid variable v, with v1 and v2 the variables to be compared between the two simulations. Then the discrepancy is dened as
R

|v1 v2 |dxdt , |v |dxdt R 1

(3)

where v1 refers to the ideal simulation. The twice shocked heavy uid is nearly stationary, so that the velocities are small and the denominator in (3) is small. Thus the large relative error in Table I for this variable reects relatively large dierences between small quantities. From Table I and Fig. 5, we conclude that our strategy of forcing a DNS simulation by an articial increase of the viscosity does not seriously degrade the simulation accuracy with regard to the macro variables. Its eect could be compared to computing with a somewhat coarser mesh, and since the mesh is already rather ne, the resulting solutions appear to be acceptable as far as macro variable convergence is concerned. We regard the modied transport parameters as turbulent viscosity coecients in a Smagorinsky type LES subgrid model. In this sense, we describe these simulations as converged LES or as DNS relative to the modied transport parameters. The benet is a possibly converged study of the micro variables.

11
Re Once shocked pure phase heavy density Twice shocked pure phase heavy density Once shocked pure phase light density Twice shocked pure phase light density Mixed phase density Once shocked pure phase heavy velocity Twice shocked pure phase heavy velocity Once shocked pure phase light velocity Twice shocked pure phase light velocity Mixed phase velocity Shock position 0.17 0.20 0.06 0.07 0.08 0.02 0.03 0.04 0.003 0.03 0.04 0.002 0.09 0.03 0.004 0.0008 0.71 0.39 0.26 0.33 0.02 0.005 0.001 0.0005 0.03 0.06 0.01 0.03 0.008 0.02 0.007 0.02 0.02 0.008 0.002 0.0007 0.02 0.01 0.01 0.01 0.005 0.001 0.0003 0.0001 5 102 3 103 3 104 4 105

TABLE I: Relative discrepancy: ne grid Sc = 10 solutions compared to an ideal simulation with zero transport, for a series of error measures. B. Convergence of the Interface Length

A second objective of our simulations, to be addressed in Sec. III, is to investigate ne scale level mixing properties, including mass fraction concentration pdfs for the distinct

12 species of the mixtures. In preparation for Sec. III, we study here statistical measures related to convergence of the interface. The basic result is that DNS, with K x (Kmesh 1), show convergence of statistics measuring ne scale interface properties. Our rst result concerns the mesh length fraction. The ratio, as dened by Fig. 3, right frame, for a xed time t = 90, is plotted in Fig. 6 for a range of simulations with a large range of viscosities and mass diusivities. The ratio is largely independent of the Schmidt number and is also mainly independent of Reynolds number and the Kolmogorov scale for Kmesh 0.5, namely for the under resolved DNS range. For these simulations, approximately 30% of the mesh blocks have a unit of interface length. This statement is independent of mesh for the range of simulations explored here. With more mesh resolution, i.e. larger values of Kmesh , the simulations are in transition to a DNS. In this case the chaotic development of the interface appears to be limited, in that mesh renement with no other change of parameters (which will increase Kmesh , decrease Remesh but not inuence Re) also decreases the mesh length fraction. The mesh related scaling of Fig. 6 is useful for understanding the under resolved DNS range of simulations, where the simulations are chaotic and growing new ne scale features at each level of mesh renement. The data collapse onto a single curve. For the DNS regime, however, it is more convenient to use physical units, with mesh dependence and convergence clearly displayed. For this purpose, we replot the same data in Fig. 7. With this scaling, we can observe the convergence of the interface length for the DNS range of simulations. Convergence occurs more rapidly for smaller Schmidt numbers and for smaller Reynolds numbers. For Sc = 0.1 and the nest mesh shown, convergence appears to occur at Reynolds numbers near 4000.

C.

An Interfacial Correlation Length Scale

We introduce a correlation length scale C to characterize the microstructure of mix. The correlation length is dened in terms of the probability of exit distance from a given phase or mean distance to the complementary phase, introduced19,20 for models of

13

0.5 0.4 0.3 0.2 0.1 0 10-3

Sc = 10 t = 90

0.5
x
200x400 400x800 800x1600 1600x3200

0.5
x
Sc = 1 t = 90
200x400 400x800 800x1600 1600x3200

Interface Fraction

Interface Fraction

x x

xx

xx

x x x

0.3 0.2 0.1 0 10-3

xx

x x

Interface Fraction

0.4

0.4 0.3 0.2 0.1 0 10-3

Sc = 0.1 t = 90

200x400 400x800 800x1600 1600x3200

xx

xx x

xx

xx

x x

10-2

Kmesh

10-1

100

101

10-2

Kmesh

10-1

100

101

10-2

Kmesh

10-1

100

101

FIG. 6: Plot of mesh length fraction vs. Kmesh for several grid levels, at t = 90. Left to right: Sc = 10, 1.0, 0.1.

Interface Fraction

Interface Fraction

20 15 10 5

15 10 5
x x x
1

Interface Fraction

Sc = 10 t = 90

200x400 400x800 800x1600 1600x3200

20

Sc = 1 t = 90

200x400 400x800 800x1600 1600x3200

20 15 10 5

Sc = 0.1 t = 90

200x400 400x800 800x1600 1600x3200

x x
1

x x

x x

x x

x x

x x

x x

x
2

x x

x x

x x

0 0 10

10

10

10

10

10

10

0 0 10

10

10

10

10

10

10

0 0 10

10

10

10

10

105

106

Re

Re

Re

FIG. 7: Plot of interface length to mixing zone area ratio (physical units) vs. Re for several grid levels, at t = 90. Left to right: Sc = 10, 1.0, 0.1.

opacity, and studied9,12 as a measure of ne scale mixing length. Using this probability measure d, we dene C = d. We plot in Fig. 8 the probability of a minimum distance for change of uid component, starting at a random point in the light uid at the mid radius, for three simulations. The exit probability data is collected into bins, each holding the data for an interval of possible exit lengths. The binned data is t to an exponential (modeling an exponential process or a Markov random eld), and from the exponent, the correlation length may be determined. Due to mathematical properties of the exponential distribution, this length is determined instead directly from the data as a correlation length. See Fig. 9.

14

10

10

10

probability density

probability density

100

100

probability density length

100

10-1

10-1

10-1

10

-2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

10

-2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

10

-2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

length

length

FIG. 8: Probability of minimum exit distance from the light phase, starting at a random light uid point on the mid radius. The dashed line in each frame is the plot of the best t exponential distribution. Three DNS, left to right, Sc = 10, 1, 0.1.
12 10 8
400x800 Sc=0.1 400x800 Sc=1 400x800 Sc=10 800x1600 Sc=0.1 800x1600 Sc=1 800x1600 Sc=10 1600x3200 Sc=0.1 1600x3200 Sc=1 1600x3200 Sc=10

0.35 0.3 0.25 0.2

400x800 Sc=0.1 400x800 Sc=1 400x800 Sc=10 800x1600 Sc=0.1 800x1600 Sc=1 800x1600 Sc=10 1600x3200 Sc=0.1 1600x3200 Sc=1 1600x3200 Sc=10

Cmesh

6 4 2 0 -1 10
0 1 2

C
0.15 0.1 0.05
4

10

10

Remesh

10

10

10

0 2 10

10

Re

10

10

10

FIG. 9: Left: plot of C /x vs. Remesh for a range of transport parameters and mesh levels, at t = 90. Right: the same data replotted as C vs. Re.

We can now assess interface convergence in terms of the behavior of C . For the under resolved regime, we look at the large Remesh limit of the left frame of Fig. 9. In this regime, Cmesh has a nearly constant value (about 2), independent of mesh and Schmidt number. As noted earlier, the nest grids display some loss of grid level complexity. For the DNS regime, we look at the smaller Reynolds numbers in the right frame of Fig. 9. This limit also suggests convergence, to an Sc dependent limit, as we expect. C can be thought of as a mean bubble or droplet radius for two phase ow having

15 volume fractions near 0 or 1 and thus having well dened continuous and separated phases. However, the denition of C does not assume a continuous and a separated phase. It is applicable to all values of the volume fraction. With our present usage, dened using the mid radius of the mixing zone, the volume fraction is near 0.5. C provides a length scale over which thermal or mass diusion must operate to smooth out or eliminate thermal or concentration uctuations. We analyze in more detail the hypothesis that the exit probability can be modeled by a exponential distribution. Having determined a correlation length for each simulation, each radius and each time, we compare the observed binned exit probabilities to a binned exponential distribution with the same correlation length. The two distributions are dened by weights w that modify the probability of a bin, relative to the uniform probability (1/number of bins) of each bin. The error in the model is then dened as |wmodel wdata |wdata / wdata , (4)

with the sum running over the bins. The exponential model error, as dened by (4) is plotted in Fig. 10 for a range of transport coecients and grids.

III.

MASS DIFFUSION AND THE SCHMIDT NUMBER

Here we study the pdf for the mass concentration of the uid mixture, dened as a function of time and radius. The mass fraction pdf depends sensitively on details of both physical and numerical modeling. To create the pdfs from the simulation data, we collect the mixture fraction variable along a constant radius at the middle of the mixing zone. Mixed cells are not averaged, but each cell fraction contributes its own mixture fraction with its own probabilities (proportional to area). The mixture fractions are then binned into 10 bins, and the resulting histogram is t to a smooth curve. We nd only a limited range of histograms from the many dierent simulations. We introduce the time dependent diusion length scale D = 2((t t0 ))1/2 , where t0 is the time of reshock. The ratio D /C is a measure of the microscopic mixing level. We plot

16
1

exponential model error

0.8 0.6 0.4 0.2 0 -1 10

400x800 Sc=0.1 400x800 Sc=1 400x800 Sc=10 800x1600 Sc=0.1 800x1600 Sc=1 800x1600 Sc=10 1600x3200 Sc=0.1 1600x3200 Sc=1 1600x3200 Sc=10

10

10

Remesh

10

10

10

FIG. 10: Plot of the exponential model error for a range of transport parameters and mesh levels, at t = 90.
12 10 8
400x800 Sc=10 800x1600 Sc=10 1600x3200 Sc=10

12 10 8
400x800 Sc=1 800x1600 Sc=1 1600x3200 Sc=1

12 10 8
400x800 Sc=0.1 800x1600 Sc=0.1 1600x3200 Sc=0.1

D/C

D/C

4 2 0 -2 -1 10 10
0

4 2 0

D/C
0 1 2 3 4

6 4 2 0

10

Remesh

10

10

10

-2 -1 10

10

10

Remesh

10

10

10

-2 -1 10

10

10

Remesh

10

10

10

FIG. 11: The ratio D /C vs. Remesh for several mesh levels and for Sc = 10, 1.0, 0.1.

this ratio vs. Remesh in Fig. 11 for a variety of Schmidt numbers and meshes. To analyze this gure, we discuss separately the small and large Remesh asymptotes. The latter is the under resolved regime. In this regime, D /C is independent of the Schmidt number, and has a value below 1, indicating little opportunity for mass diusion to occur. In contrast, the small Remesh limit in this gure represents the DNS. These are strongly dependent on Sc. They appear to be converged to a grid independent value.

17
6 6 5

Probability density

Probability density

4 3 2 1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

4 3 2 1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Probability density

= 0.11 D /C = 0.98

= 0.23 D /C = 3.29

4 3 2 1

= 0.38 D /C = 5.25

0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Light fluid mixture fraction

Light fluid mixture fraction

Light fluid mixture fraction

FIG. 12: Pdfs of light species volume fraction at time t = 90. The data has been collected into 10 bins; we show both the resulting histogram and a smooth (Akima spline) interpolation. Left: typical bimodal case (Kmesh = 0.7, Sc = 10). Center and right: increasingly diused light uid cases (Kmesh = 1.1, Sc = 1.0 and (Kmesh = 0.8, Sc = 0.1, respectively). The mesh is 1600 800 in all cases.

We anticipate that the ratio D /C largely determines the mass concentration pdfs. In other words, the pdfs are naturally approximated as belonging to a one parameter family, indexed by D /C . We return to this important point later. Most of the pdfs are strongly bimodal, with narrow peaks at mixture fractions perhaps 0.1 and 0.9. For a few, the light uid peak (at 0.9) is considerably diused or even missing completely. We display typical plots of the pdfs representative of these types. All of the simulations in the under resolved region (Kmesh 0.5) are strongly bimodal. In the DNS regime (Kmesh 1), the highly diusive simulations (Sc = 0.1) have a partially or highly diused light uid component, while the less diusive simulations (Sc 1 and especially Sc = 10) do not. See Fig. 12. The trends in the pdfs are as predicted by the trends in the D /C ratio. The data presented is from the time t = 90, which is the beginning of the chaotic stage of interface development. We also examined the light uid mixture fraction pdf at t = 110, in order to allow more time for diusion, and did not observe a striking change in the pdfs. All the simulations of Fig. 12 are in the DNS regime, with Kmesh = O(1) and some degree of convergence for the interface length. Thus the progression of increased diusion

18 observed in Fig. 12 reects a change in the physical formulation of the problem, namely an approximate 10 fold increase in mass diusion with each successive frame, and is not related to mesh convergence or subgrid modeling issues. To understand the greater sensitivity of the light uid peak than the heavy peak to mass diusion, we note that the result of diusing heavy uid into a region of light uid can easily make a large change in the mass weighted mixture fraction, while the diusion of the same fraction of light uid into a region of heavy uid will have a relatively smaller eect. The mean molecular mixing fraction , dened25 as = f1 f2 f1 f2 (5)

is a common measure of mixing. Here fk is the mass concentration fraction for the species k. With Sc = , obviously there will be zero mass diusion, and = 0. The probability distribution function for the species concentrations fi is a more rened measure of mixing, and provides the level of detail needed for prediction of turbulent combustion. In each frame of Fig. 12, the values of = (rmid ) on the mid radius of the mixing zone and D /C are indicated. depends sensitively on the grid resolution for the under resolved regime. In the DNS regime it depends on the transport coecients but should be mesh independent. The values of we report are smaller than those commonly reported for simulation studies of mixing, especially for Rayleigh-Taylor mixing.

IV.

A MATHEMATICAL AND COMPUTATIONAL FOOTNOTE A. Mathematical Existence and Uniqueness Theories

If it is true that numerical solutions to the Euler equations are underdetermined, and that the convergence process produces new structures on each length scale, what is the consequence of these facts for a mathematical existence theory for these equations? It would appear that if such a mathematical solution were to exist, it might not be

19 a function, but a generalized function in the sense of compensated compactness. The compensated compactness generalized functions are pdfs depending on space and time, not dissimilar from what is observed computationally. In the one dimensional existence theory of compensated compactness of DiPerna16,17 and Ding and co-workers3,5 , the rst and easier step is to show existence of the generalized solution. The more dicult second step is to show that a generalized solution is a classical (weak) one. In view of the slow rate of progress with the existence theory in two dimensions, and in view of the possibility that this second step might not be correct, it is worth trying to establish generalized compensated compactness solutions without the step of passing to a classical one. Typically such proofs are based on compactness, a method of proof which yields existence but not uniqueness for the weak solutions. We have already commented on the possible nonuniqueness of solutions to the Euler equations. We also note that weak solutions in the form of a space time dependent pdf for the primitive variables are actually what is required scientically for a combustion simulation. For this reason, such pdf weak solutions couple into the requirements of computational physics, whether they are required as a fundamental scientic truth or not.

B.

A Practical LES Subgrid Model

We also speculate on the consequences of the ideas expressed here for practical simulations of uid mixing. We start by recalling the two types of objectives for the simulations. The rst, which we call macro objectives, are to obtain locations of the principal waves and mean ow quantities in agreement with a ne grid simulation having minimal or zero transport coecients. The comparison simulations have viscosity and mass diusion coecients set to the molecular values or as modied by turbulent subgrid models. The second objective is to obtain converged values for the joint pdfs of temperature and mass concentration of the mixing species. Since these pdfs appear to depend on the physical choice of the transport coecients, the convergence algorithm must be parameterized to be consistent with the physics. We call this the micro objective. Both are required

20 within a single simulation. To attain these two goals, and a third goal of practical or feasible levels of simulation, we are proposing two dierent classes of simulations. The purpose of the rst is to obtain, to the extent possible, direct knowledge of the micro observables. Here we allow articially high levels of viscosity, as long as the macro variables are in approximate agreement with the reference solution. For such simulations, the interface appears to have converged, and we use molecular values of the Schmidt number to set values of the coecient of mass diusion. Continued mesh renement should lead to stable values for the micro variables. We anticipate a one parameter family of concentration pdfs as the Schmidt number or the ratio D /C is varied. This is a conjecture which has still be veried. If we manage in this manner to obtain converged values for the micro observables, we then move to our second class of simulations. For this class of simulations, we set the physical viscosity to its molecular value, perhaps as modied by a subgrid model to account for the inuence of the unresolved scales on those being computed. With the knowledge of the concentration pdfs from the rst class of simulations, we can also adjust the mass diusion coecient in the second class of simulations to give micro observables in agreement with the rst. To achieve this agreement, we depend on the conjecture that the concentration pdfs lie approximately on a one parameter family, which can be indexed by the mass diusion (i.e., the Schmidt number or the ratio D /C ). Here the grids can be notable coarser, and the simulations should be readily feasible. This program appears to be consistent with accepted ideas in turbulence modeling18 . In future work, we hope to ll in the gaps in this program and establish the practicality of simulations which are converged in both their micro and macro observables.

This work was supported in part by U.S. Department of Energy grants DE-AC02-98CH10886 and DE-FG52-06NA26208, and the Army Research Oce grant W911NF0510413. The sim-

21
ulations 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. Los Alamos National Laboratory Preprint LA-UR-08-0068.
1

G. Batchelor. The Theory of Homogeneous Turbulence. The University Press, Cambridge, 1955.

William Cabot and Andrew Cook. Reynolds number eects on Rayleigh-Taylor instability with possible implications for type Ia supernovae. Nature Physics, 2:562568, 2006.

G.-Q. Chen. Convergence of the Lax-Friedrichs scheme for isentropic gas dynamics III. Acta Mathematica Scienitia, 6:75120, 1986.

P. Colella. A direct Eulerian MUSCL scheme for gas dynamics. SIAM Journal on Computing, 6(1):104117, 1985.

X. Ding, G.-Q. Chen, and P. Luo. Convergence of the Lax-Friedrichs scheme for isentropic gas dynamics I and II. Acta Mathematica Scientia, 5:415432, 433472, 1985.

Jian Du, Brian Fix, James Glimm, Xicheng Jia, Xiaolin Li, Yunhua Li, and Lingling Wu. A simple package for front tracking. J. Comp. Phys., 213:613628, 2006. Stony Brook University preprint SUNYSB-AMS-05-02.

S. Dutta, E. George, J. Glimm, J. Grove, H. Jin, T. Lee, X. Li, D. H. Sharp, K. Ye, Y. Yu, Y. Zhang, and M. Zhao. Shock wave interactions in spherical and perturbed spherical geometries. Nonlinear Analysis, 63:644652, 2005. University at Stony Brook preprint number SB-AMS-04-09 and LANL report No. LA-UR-04-2989.

J. Glimm, J. W. Grove, Y. Kang, T. Lee, X. Li, D. H. Sharp, Y. Yu, K. Ye, and M. Zhao. Statistical Riemann problems and a composition law for errors in numerical solutions of shock physics problems. SISC, 26:666697, 2004. University at Stony Brook Preprint Number SBAMS-03-11, Los Alamos National Laboratory number LA-UR-03-2921.

J. Glimm, J. W. Grove, X. L. Li, W. Oh, and D. H. Sharp. A critical analysis of RayleighTaylor growth rates. J. Comp. Phys., 169:652677, 2001.

22
10

D. J. Hill, C. Pantano, and D. L. Pullin. Large-eddy simulation and multiscale modeling of a Richtmyer-Meshkov instability with reshock. J. Fluid Mech, 557:2961, 2006.

11

H. Lee, H. Jin, Y. Yu, and J. Glimm. On the validation of turbulent mixing simulations of Rayleigh-Taylor mixing. Phys. Fluids, 2007. Accepted for publication. Stony Brook University Preprint SUNYSB-AMS-07-03.

12

H. Lim, Y. Yu, H. Jin, D. Kim, H. Lee, J. Glimm, X.-L. Li, and D. H. Sharp. Multi scale models for uid mixing. Special issue CMAME, 2007. Accepted for publication. Stony Brook University Preprint SUNYSB-AMS-07-05.

13

X. F. Liu, Y. H. Li, J. Glimm, and X. L. Li. A front tracking algorithm for limited mass diusion. J. of Comp. Phys., 2007. Accepted. Stony Brook University preprint number SUNYSB-AMS-06-01.

14

T. O. Masser. Breaking Temperature Equilibrium in Mixed Cell Hydrodynamics. Ph.d. thesis, State University of New York at Stony Brook, 2007.

15

A. S. Monin and A. M. Yaglom. Statistical Fluid Mechanics: Mechanics of Turbulence. MIT Press, 1971.

16

R. Di Perna. Convergence of approximate solutions to conservation laws. Arch. Rational Mech. Anal., 82:2770, 1983.

17

R. Di Perna. Compensated compactness and general systems of conservation laws. Trans. Amer. Math. Soc., 292:383420, 1985.

18

Heintz Pitsch. Large-eddy simulation of turbulent combustion. Annual Rev. Fluid Mech., 38:453482, 2006.

19

G. C. Pomraning. Linear kinetic theory and particle transport in stochastic mixtures, volume 7 of Series on Advances in Mathematics for Applied Sciences. World Scientic, Singapore, 1991.

20

G. C. Pomraning. Transport theory in discrete stochastic mixtures. Advances in Nuclear Science and Technology, 24:4793, 1996.

21

D. Pullin. A vortex based model for the subgrid ux of a passive scalar. Phys. of Fluids,

23
12:23112319, 2000.
22

D. Pullin and T. S. Lundgren. Axial motion and scalar transport in stretched spiral vortices. Phys. of Fluids, 13:25532563, 2001.

23 24

D. H. Sharp. An overview of Rayleigh-Taylor instability. Physica D, 12:318, 1984. P. Woodward and P. Colella. Numerical simulation of two-dimensional uid ows with strong shocks. J. Comp. Phys., 54:115, 1984.

25

D. L. Youngs. Three-dimensional numerical simulation of turbulent mixing by RayleighTaylor instability. Phys. Fluids A, 3:13121319, 1991.

26

Y. Yu, M. Zhao, T. Lee, N. Pestieau, W. Bo, J. Glimm, and J. W. Grove. Uncertainty quantication for chaotic computational uid dynamics. J. Comp. Phys., 217:200216, 2006. Stony Brook Preprint number SB-AMS-05-16 and LANL preprint number LA-UR-05-6212.