Out 3

Numerical modeling in turbulent mixing flows
A Dissertation Presented
by
Hyunkyung Lim
to
The Graduate School in Partial Fulfillment of the Requirements for the
Degree of
Doctor of Philosophy
in
Applied Mathematics and Statistics

Stony Brook University
August 2009
UMI Number: 3401712
All rights reserved

INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.
In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.
UMI 3401712
Copyright 2010 by ProQuest LLC.
All rights reserved. This edition of the work is protected against
unauthorized copying under Title 17, United States Code.
ProQuest LLC
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106-1346

The Graduate School
Hyunkyung Lim
We, the dissertation committee for the above candidate for the Doctor of
Philosophy degree, hereby recommend acceptance of this dissertation.
James Glimm
Advisor
Department of Applied Mathematics and Statistics
Xiaolin Li
Chairman
Xiangmin Jiao
Member
Foluso Ladeinde
Outside Member
Department of Mechanical Engineering
This dissertation is accepted by the Graduate School.

Lawrence Martin
Dean of the Graduate School
ii
Abstract of the Dissertation

Numerical modeling in turbulent mixing flows
by
Hyunkyung Lim
Doctor of Philosophy
in
Applied Mathematics and Statistics
2009
We are concerned with the chaotic flow fields of turbulent mixing. Chaotic
flow is found in an extreme form in multiply shocked Richtmyer-Meshkov unstable flows. The goal of a converged simulation for this problem is twofold:
to obtain converged solutions for macro solution features, such as the trajectories of the principal shock waves, mixing zone edges, and mean densities and
velocities within each phase, and also for such micro solution features as the
joint probability distributions of the temperature and species concentration or
a chemical reaction rate. We introduce parameterized subgrid models of mass
and thermal diffusion, to define the large eddy simulation (LES) that replicate
the micro features observed in the direct numerical simulation (DNS). The
Schmidt numbers and Prandtl numbers are chosen to represent typical liquid,
iii
gas and plasma parameter values. Our main result is to explore the variation
of the Schmidt, Prandtl and Reynolds numbers by three orders of magnitude,
and the mesh by a factor of 8 per linear dimension (up to 3200 cells per dimension), to allow exploration of both DNS and LES regimes and verification
of the simulations for both macro and micro observables. We study mesh convergence for key properties describing the molecular level of mixing, including
chemical reaction rates between the distinct fluid species.
Methodologically, the results are also new. In common with the shock
capturing community, we allow and maintain sharp solution gradients, and
we enhance these gradients through use of front tracking. In common with
the turbulence modeling community, we include subgrid scale models with
no adjustable parameters for LES. These two methodologies have not been
previously combined. In contrast to both of these methodologies, our use of
Front Tracking, with DNS or LES resolution of the momentum equation at or
near the Kolmogorov scale, but without resolving the Batchelor scale, allows
a feasible approach to the modeling of high Schmidt number flows.
Key Words: Turbulence, Subgrid models, Large eddy simulation, Direct
numerical simulation, Mass diffusion, Thermal diffusion, Schmidt numbers,
Prandtl numbers
iv
To my parents
Table of Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
xii
List of Tables
Acknowledgements
. . . . . . . . . . . . . . . . . . . . . . . . . xiii
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1
Overview and Motivation . . . . . . . . . . . . . . . . . . . . .
1.2
Front Tracking Method . . . . . . . . . . . . . . . . . . . . . .
1.2.1
Front Tracking Algorithm . . . . . . . . . . . . . . . .
1.2.2
Conservative Front Tracking Algorithm . . . . . . . . .
Dissertation Organization . . . . . . . . . . . . . . . . . . . .
11
. . . . . . .
12
2.1
The Governing Equations . . . . . . . . . . . . . . . . . . . .
12
2.2
Numerical Development for LES . . . . . . . . . . . . . . . . .
16
1 Introduction
1.3
2 Developments for LES with turbulence models
2.2.1
A Subgrid Scale Dynamic Model for the Momentum

Equation . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2
18
A Subgrid Scale Dynamic Model for the Energy Equation 23

vi
2.2.3
A Subgrid Scale Dynamic Model for the Concentration

Equation . . . . . . . . . . . . . . . . . . . . . . . . . .
25
. . . . . . . . . . . . . . . . . .
27
3.1
Test Problem Setup . . . . . . . . . . . . . . . . . . . . . . . .
28
3.2
Research Objectives . . . . . . . . . . . . . . . . . . . . . . . .
32
. . . . . . . . . .
40
Correlation and Diffusion Length Scales . . . . . . . . . . . . .
42
. . . . . . . . . . . . . . .
46
5.1
Joint PDFs . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.2
Concentration Moments . . . . . . . . . . . . . . . . . . . . .
49
5.3
Temperature PDF . . . . . . . . . . . . . . . . . . . . . . . .
52
5.4
Chemical Production . . . . . . . . . . . . . . . . . . . . . . .
55
5.5
Statistical Fluctuation vs. Mesh Convergence . . . . . . . . .
56
. . . . .
58
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
3 The Flow Instability Problem
4 Mesh Convergence for Macro Observables

4.1
5 Convergence of Micro Observables
6 Mathematical Existence and Uniqueness Theories

7 Conclusion
vii
List of Figures
1.1
A sample of three types of interface representations. Upper-left:

grid free (GF) interface; Upper-right: grid based (GB) interface;
Lower: locally grid based (LGB) interface. . . . . . . . . . . . .
3.1
Initial (left) and late time (right) density plot for the RichtmyerMeshkov fluid instability (unregularized, with zero physical transport coefficients) under study in this paper. . . . . . . . . . . .
3.2
29
Space time (r, t) contours of the primary waves, as detected by

the wave filter algorithm. These are the inward (direct) and outward (reflected) 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]. . . . . . . . . . . . . . . . .
3.3
31
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.
viii
32
3.4
Chemical reaction rate at t = 90 for three representative values for the transport parameters, and each for three values of
the Reynolds number. The activation temperature is chosen as
15, 000oK. The point of this plot is the strikingly different values
obtained for hwi for the nine physical regimes considered here.
4.1
39
Time integrated relative discrepancy in the mixing zone edge locations as compared to a fine grid, zero transport simulation.
The discrepancy, for cases l and g, is mainly due to mesh errors
and is partly due to modification of physical parameters. The
discrepancy decreases with mesh refinement, uniformly as physical parameters are varied. The plasma case p shows convergence,
but to a Reynolds number dependent limit distinct (for small Re)
from the zero transport case.
4.2
. . . . . . . . . . . . . . . . . . .
41
Left: plot of C /x vs. Kmesh for a range of mesh levels and

transport parameters cases l, g, p from Table 3.1, at t = 90,
r = rmid. Right: the same data replotted as C vs. K . . . . . .
4.3
42
The ratios (D + Dt )/C and (T + Tt )/C vs. K for several

mesh levels and for transport coefficients cases l, g, p from Table 3.1. The values to the far right in each frame are momemtumresolved, and appear to be mesh converged for the finer meshes.
5.1
44
Case l. The joint pdf of light species concentration and temperature at time t = 90. The data has been collected into 10 10
bins. The mesh is 1600 3200 and Re 6000.
ix
. . . . . . . . .
47
5.2
Case g. The joint pdf of light species concentration and temperature at time t = 90. The data has been collected into 10 10
5.3
. . . . . . . . .
48
Case p. The joint pdf of light species concentration and temperature at time t = 90. The data has been collected into 10 10
5.4
. . . . . . . . .
48
Plot of (T ) vs. T . Cases l, g, p (rows 1 to 3), with Re increasing,

left to right. Error bars indicate coarse grid ensemble fluctuation
(2) centered at the ensemble mean, for cases l, g only.
5.5
5.6
. . .
53
Plot of T pdf. Cases l, g, p (rows 1 to 3), with Re increasing, left

to right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
Radial dependence for hwi for the case l, Re 6000. . . . . . .
57
List of Tables
3.1
Transport coefficients considered in this paper, arranged in order

of increasing diffusivity. The presumed typical numerical transport parameters are shown for reference only and are not explored
here. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1
34
Ratios of dimensionless molecular to turbulent Schmidt and Prandtl

numbers for various cases, all with the finest grid. Ratios large relative to unity indicate resolution of the associated transport equation. The momentum equation has full resolution for Re 300
only. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1
43
Mesh errors for the joint temperature-concentration pdfs to illustrate possibilities of mesh convergence. Comparison is coarse (c)
to fine (f) and medium (m) to fine with pdfs compared using the
Kolmogorov-Smirnov metric. Various physical cases reported. .
5.2
49
Mesh errors for (T ), defined by the weighted L1 norm (5.4).

Comparison is coarse (c) to fine (f) and medium (m) to fine.
Various physical cases reported.
xi
. . . . . . . . . . . . . . . . .
51
5.3
Mesh errors for the specific chemical production w for an activation temperature TAC = 8, 000oK. The comparison, based on
the Kolmogorov-Smirnov matric, shows mesh convergence. The
comparison is coarse mesh (c) to fine (f) and medium (m) to fine.
Various physical cases reported.
5.4
. . . . . . . . . . . . . . . . .
52
Errors for the temperature pdfs measured in the KolmogorovSmirnov metric show mesh convergence. Comparison is coarse
mesh (c) to fine (f) and medium (m) to fine. Various physical
cases reported.
5.5
. . . . . . . . . . . . . . . . . . . . . . . . . . .
55
Relative coarse grid ensemble fluctuations (2) divided by ensemble mean, and mesh errors for the specific chemical production
w for an activation temperature TAC = 15, 000oK. Comparison
based on the Kolmogorov-Smirnov metric. Comparison is coarse
mesh (c) to fine (f) and medium (m) to fine. Only cases l and g
are reported.
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
xii
56
Acknowledgements
I would like to express my profound gratitude to my advisor, Professor

James Glimm, for suggesting this important and exciting thesis topic and for
his advice, support, encouragement and guidance toward my Ph. D. degree.
He taught me not only the way to do scientific research, but also the way to become a professional scientist. He was always enthusiastic about my education.
He is my advisor and a lifetime role model for me.
I am deeply indebted to the support of Professor Xiaolin Li. He provided
encouragement and valuable technical knowledge and helped me maintain my
focus on the research. I would also like to thank Professor Foluso Ladeinde for
being on my dissertation committee.
I would like to thank Professor Xiangmin Jiao and Roman Samulyak, Drs.
John Grove and David H. Sharp from whom I have learned many important
scientific and mathematical skills.
I would like to thank all my friends during my years of study as a graduate
student at Stony Brook for their friendship and encouragement. In particular,
I would like to mention Drs. Yan Yu, Wurigen Bo and Daesang Kim, Tulin
Kaman, Justin Iwerks, Xingtao Liu, Lingling Wu, Brian Fix. They have shared
with me many interesting and inspiring ideas.
Throughout my academic career, the constant support of my parents has

always motivated me to strive forward. Their unconditional love has never
been affected by the physical distance between us. My dissertation is dedicated
to them.
Chapter 1
Introduction
1.1
Overview and Motivation

Turbulence is a prototypical multiscale problem, with a cascade of length
scales generally too broad to be modeled effectively by explicit numerical algorithms. The dimensionless Reynolds number governs this process. Turbulent mixing couples turbulence to a concentration equation and introduces
the Schmidt number and the Prandtl number as dimensionless parameters.
We consider flows which are compressible and which couple the concentration
equation actively into the flow dynamics (due to a density contrast, measured
by the dimensionless Atwood number A = (2 1 )/(2 + 1 ), with A > 0).
These two properties introduce features into the modeling not present in theories of passive scalar transport by a turbulent field [2, 43]. Acceleration driven
turbulent mixing is a classical hydrodynamical instability, in which acceleration is directed across a fluid interface separating distinct fluids of different
densities [53]. Steady acceleration of a density discontinuity defines the classical Rayleigh-Taylor (RT) instability and impulsive acceleration produced by a
1
shock wave passing through the fluids defines the Richtmyer-Meshkov mixing.
We are concerned with the Richtmyer-Meshkov (RM) instability.
Numerical modeling of turbulence requires removal of degrees of freedom, so that the ones that remain can fit into a feasible computation. There
are three numerical models in the computation such as Reynolds averaged
Navier-Stokes simulations (RANS), large eddy simulations (LES) and direct
numerical simulations (DNS). Removal of fluctuations, to study statistically
mean quantities leads to RANS equations. Removal of small length scales, to
study the large scale flow features leads to LES, while the restriction to small
scale problems allows DNS, in which all transport effects are fully resolved.
These terms have analogous meanings in the case of turbulent mixing. Both
RANS and LES call for closure models, to introduce the influence of the omitted scales upon those computed. LES are formulated in terms of an averaging
procedure. When this average is replaced by the mesh block average over the
computational grid, the LES method is called implicit (ILES). The closure
models generally contain parameters which are set by comparison to a more
exact computation. In this sense LES and DNS (and physical experiments to
the extent available) are used to validate the parameters in RANS models and
DNS is used to validate those in an LES model.
Since we recognize at the outset that it is not possible to represent all
details of a turbulent mixing flow in a feasible computation, we list here the
types of observables we would hope to compute correctly. The most important
macroscopic variables are those which define the mixing zone, that is the edge
position for the mixing zone as a function of time, and the time dependent
2
locations of the principal shock waves (if any) which play a role in defining
it. Additional macroscopic variables which serve to define the flow include
the dominant size of the bubbles (light fluid inclusions in the heavy fluid)
and mean fluctuations of the mixing zone edge. From the point of view of
turbulent combustion, the microscopic (atomic) properties of the mixture are
important, to specify the stochiometry needed for combustion. According to
[47], the joint probability distribution of the species concentrations and of the
temperature is a required input for LES combustion models.
1.2
Front Tracking Method

Discontinuities in the solutions of nonlinear conservation laws are a pri-
mary difficulty for numerical simulations. Commonly used numerical schemes

are convergent at higher order only for smooth solutions. At discontinuities,
the local truncation errors are first order and the solutions are not convergent
in a point-wise sense. For nonlinear discontinuities, the width of the solution
error region does not grow in time, and is generally about two mesh blocks,
but for linear discontinuities, the error region is wider, some five mesh blocks,
and is generally growing in time. The front tracking method was introduced
[1719, 2628] to solve the discontinuity problem, especially for the representation of linear discontinuity such as fluid layer boundaries. The front tracking
method is initiated by Richtmyer and Morton [51] and extensive work on the
front tracking method has been carried out by J. Glimm and his coworkers
[6, 15, 18, 19, 21, 2628]. The method has been implemented and developed
in a robust and validated code called FronTier [1720]. The front tracking
3
method tracks a discontinuity interface by using analytical solutions of Riemann problems across the interface, and it applies a finite difference scheme
to solve the equation on different sides of the discontinuity interface by using
the ghost cell extrapolation method [21, 26]. This ghost cell method was also
used in tracking using the level set method [12, 13]. The ghost cell method
does not preserve conservation at the cells cut by the interface and it has less
than first order accuracy at discontinuities. The algorithm is conservative for
interior regular cells, but it is not conservative at the irregular cells near the
interface. A fully conservative front tracking algorithm has been implemented
in FronTier. The algorithm is based on the previous works [2225, 36, 37] with
dimensional generalization and algorithmic refinement. Also there are some
pioneer works on the conservative tracking algorithms by Chern and Colella
[5], Pember, Bell and Colella [44], Swartz and Wendroff [54], and the early one
by Harten and Hyman [29]. In this algorithm the front is tracked as a space
time surface (three dimensions in the four dimensional space time, for the case
of three spatial dimensions). The space time cells cut by this surface are differenced using a finite volume approach. We will give a brief review of the
front tracking and fully conservative front tracking algorithm in the FronTier
code.
1.2.1
Front Tracking Algorithm
The front tracking method in the FronTier is performed by two main

steps. The first is to propagate and update the front dynamically. The second
is to update the interior states.
4
In the interface propagation, the interface points are first propagated

normally. By computing the solution to the local Riemann problem with
initial states being those on either side of the interface point and using the
method of characteristics, a wave speed and a new position for the interface
point are determined. The interface states are then updated by a tangential
sweep, which uses a chosen interior solver with a stencil centered at the new
interface point.
After interface propagation, we get a new interface at new time level.
Any tangles in the new interface are need to be resolved. To resolve this
problem and reduce the variance in size and aspect ration of the segments
(2D) or triangles (3D) making up the interface, the new interface need to be
reconstructed. There are three types of interface reconstruction such as the
grid based (GB) interface tracking [39], the grid free (GF) interface tracking
[15, 18, 19] and the locally grid based (LGB) interface tracking. Fig. 1.1 shows
examples of these three types of interface reconstruction in 2D. A grid free
interface is independent of the underlying grid. By reconstruction, all the
points on a grid based interface lie on the cell edges. Thus, the interface
elements are constructed from vertices which are the intersections between the
interface and the grid lines. The grid free interface uses the locally grid based
(LGB) method [10] to resolve topological bifurcations. The LGB method uses
the GF interface except at a region of space and time where interface entities
intersect with each other (or with itself). When such a topological bifurcation
occurs, we first isolate the troubled region using a minimized but sufficient
box to contain the bifurcation. We then use the GB method to reconstruct
5
the interface inside the box. Finally, we relink the interface segment inside
the box with the ambient grid free interface. The LGB method combines the
robustness of the grid based method with the accuracy of grid free method. It
is thus a significant improvement to both of these algorithms.
As the next step, we update all the interior states on cell centers. A
connected region in the domain separated by the interface is represented by a
component. Therefore, each grid node is associated with a specific component
in addition to the state variables. The interior states are updated by finite
difference schemes.
1.2.2
Conservative Front Tracking Algorithm
Consider a system of hyperbolic conservation laws in N spatial dimension

in differential form
U
+ F (U) = 0, F = (f1 , f2 , ..., fN ) ,
t
(1.1)
where U Rp and fj (U) = (f1j (U), ..., fpj (U))T Rp are defined in a spatial
domain RN .
Integrating (1.1) in a time-space domain V RN +1 , we obtain the integral form of (1.1),
U
+ F (U))dV = 0 .
t
(1.2)
By the divergence theorem, we have

Z
(U, F (U)) ndS = 0 .
(1.3)
Figure 1.1: A sample of three types of interface representations. Upper-left:

grid free (GF) interface; Upper-right: grid based (GB) interface; Lower: locally
grid based (LGB) interface.
The finite difference method presented here is an explicit finite volume integration scheme based on the integral form (1.3).
Assume a space-time discretization {Vi } which conforms to the spacetime interface as U evolves in one time step from tn to tn+1 . We solve (1.3)
explicitly in this region. We define each Vi as a space-time control volume, and
Vi = Din Din+1 Si with Din , Din+1 , and Si meet at most at their boundaries,
where Din and Din+1 are the boundary surfaces of Vi at time level n and n + 1
respectively, and Si is the complementary boundary surface of Vi . Dividing
the calculation of the integral (1.3) into three parts over Din , Din+1 and Si
respectively, we have
|Din+1|U |tn+1 =
where
1
U |tn = n
|Di |
and
U |tn+1 =
1
|Din+1 |
|Din |U |tn
Si
(U, F (U)) ndS ,
(1.4)
U(x1 , ..., xN , tn )dx1 ...dxN
|Din |
|Din+1 |
U(x1 , ..., xN , tn+1 )dx1 ...dxN
are cell averages, |Din | is the face area of Din , |Din+1 | is the face area of Din+1 ,
and n is the outward normal of Si . Therefore, U |tn+1 , is the solution to (1.4)
at time tn+1 .
To calculate U |tn+1 , we first need to determine the space-time control
volume {Vi }. Then we calculate the fluxes defined on the surfaces of Vi , so
that we can apply (1.4).
8
Eq. (1.4) degenerates into a static finite volume scheme
U |tn+1 = U |tn
Si
F (U) ndS ,
(1.5)
when the control volume has a constant face shape Din+1 = Din . In such a
case, the complementary boundary surface Si is parallel to the time axis.
The discontinuities of the numerical solution u form an INTERFACE
(set of POINTs for a 1D ambient dimensional space, CURVEs for 2D, and
SURFACEs for 3D), which is propagated from one time level to the next.
These interface constituents are described in a piecewise linear manner, and
made up of linear elements (simplexes), which are defined by their vertices, the
interface POINTs. The front POINTs are propagated by solving the Riemann
Problem in the normal direction and a tangential sweep to update the states
on the interface. Here and below, we use the dimension D of an interface
to refer to the ambient space in which it is embedded. Thus D may be the
dimension of space or of space time, and the elements of the D dimensional
interface have intrinsic dimension D - 1 (co-dimension 1) or lower. See [1820]
for details.
The conservative front tracking algorithm contains the following major
steps:
1. To propagate the spatial interface at time level n to a new interface at time
level n + 1. The grid-free interface, or its LGB reconstruction in the
case of spatial tangles, is the permanent interface used within the time
step propagation loop, while the space-time GB interface is only used for
9
the space-time control volume reconstruction. While the clipping of the

grid-free interface to become grid-based introduces a truncation error at
each time step, this error is not involved in the interface propagation to
subsequent time steps.
2. To construct the space-time interface by joining two spatial interfaces at
time level n and n + 1. Given two piecewise linear spatial INTERFACEs
which are separated in time by a step t, we construct (triangulate) a
space-time surface joining the two spatial INTERFACEs. To construct
the space-time interface we perform the following two major tasks for
every space-time cell: (1) Construct a grid-free space-time interface by
joining the interface at time step n and the new propagated grid free
interface at time step n + 1; (2) Reconstruct the control volume by
the convex hull method using the vertices of the grid based space-time
interface and corners of the regular space-time control volume.
3. To construct space-time control volumes by using the intersection points
of the space-time interface and the space-time cell edges. Those control
volumes with small or no top will be merged with their neighboring
volumes to satisfy the CFL condition.
4. To use a higher order Godunov type finite volume scheme as in to update
the states.
By joining the spatial interfaces at time level n and n + 1, we construct a
space-time manifold which is called the space-time interface. The space-time
interface cuts the space-time grid cells into fragments when the grid cell and the
10
interface intersect. These fragments together with the regular space time grid
cells which do not meet the interface are the space-time control volumes. We
apply the finite volume scheme to them to calculate the numerical solutions.
The LGB (GF when there is no topological bifurcation) propagated interface poses a difficult problem to the construction of irregular control volume.
Therefore, to simplify the procedure of constructing space-time control volume {Vi } in (1.4), we clip the GF space-time interface into a GB space-time
interface. As a result, the space-time control volumes are grid based and their
vertices are intersections of the space-time interface and regular space-time
cell edges.
1.3
Dissertation Organization
Chapter 1 first gives the overview and brief review of the front tracking
and fully conservative front tracking method.

Chapter 2 present the governing equation and the numerical development
for LES with turbulence models.
In Chapter 3 we describe the test problem setup and our goal. We explore
mesh refinement and mesh convergence and dependence on a range of Reynolds
numbers.
In Chapter 4 and 5, we present microscopic and macroscopic aspects of
mixing. Specializing to a 2D circular RM flow, we find sensitivities to numerical and physical modeling for macroscopic variables and microscopic variables.
The interface length depends sensitively on the numerical and physical modeling. Consequently, interface sensitive quantities such as mass diffusion and
11
atomic level mixing have a similar dependence. We also observe that macroscopic variables are insensitive to variation of physical and numerical modeling
while the microscopic variables are highly sensitive.
Chapter 6 discusses the mathematical existence and uniqueness theories
and the conclusion is discussed in Chapter 7.
12
Chapter 2
Developments for LES with turbulence models
The governing equation and the numerical development for LES are presented in this chapter.
2.1
The Governing Equations

We study the compressible Navier-Stokes equations with viscosity, mass
diffusion and thermal conductivity, for two miscible species initially separated
by a sharp interface. The primitive equations describe the DNS limit, in
which transport effects are resolved. A measure of this limit, as applied to the
momentum equation, is the criteria Kmesh 1 where Kmesh = K /x and
K is the Kolmogorov length scale,
K = (k3 /)1/4 ,
(2.1)
= k |S|2 ,
(2.2)
where
13
S is the strain rate tensor

1
Sij =
2
vi
vj
+
xj
xi
(2.3)
defined in terms of the velocity v and for any matrix A = Aij ,

|A|2 =
2A2ij .
(2.4)
A related convergence measure, but applicable to the concentration and

energy equations is that the ratio of the turbulent transport parameters (i.e.
the coefficients of the dynamic subgrid scale models (given below)) to the
molecular ones be small.
LES start from a filter, or averaging procedure, applied to the primitive
equations of compressible flow. We adopt what is known as an implicit filter,
namely a grid block average. In this case the quantities in the defining equations are averaged over a grid block. New terms, arising from the average of
the nonlinear terms, are introduced into the equations. We use a conventional
definition of these terms, following refs.[14, 40, 42]. The subgrid models are parameterized dynamically, meaning that the model parameters are determined
completely from the resolved scales. In this sense, the models are parameter
free. For DNS, these terms have little effect.
We write the filtered continuity, momentum, energy and concentration
equations for two miscible fluid species in an inertial frame. The filtered quantities are considered to be mesh block averages, and denoted with an overbar,
while mass averaged quantities are denoted with a tilde. Repeated indices are
14
summed.
vei
=0,
+
t
xi
dij ij
vej (vei vej + pij )
=
,
+
t
xi
xi
xi
E (E + p)vei
dij vej
=
+
+
t
xi
xi
xi
(H)
(T )
(V )
Te
xi
(2.5)
xi
fh H
fl )D
e
(H
1 kk vei qi
q
q
+
i i
,
2 xi
xi
xi
xi
!
()
e
e evei
e
+
D
i ,
=
t
xi
xi
xi
xi
e
xi
(2.6)
!
(2.7)
(2.8)
(2.9)
(H)
where the subgrid scale (SGS) variables are the ij , qi
(T )
(V )
, qi , qi
()
and qi .
They are expressed as
ij = (g
vi vj vei vej )
(H)
qi
(T )
qi
(V )
qi
()
qi
e ei )
= (c]
p T vi cep T v
1
= (v^
ei )
k vk vi vek vek v
2
= dij vj dij vej
fi evei ) .
= (v
(2.10)
(2.11)
(2.12)
(2.13)
(2.14)
e vei , p and E denote, respectively,

The dependent filtered variables , ,
the total mass density, the species mass fraction, the velocity, the pressure,
15
and the total specific energy with

E = e
e + vek 2 /2 + kk /2 .
(2.15)
fh and H
fl are the partial specific enthalpy of each species defined by
Here H
fh = eeh + p
H
(2.16)
fl = eel + p ,
H
(2.17)
where eeh and eel are the specific internal energy of each species. The equation
of state for each of the species is taken to be a stiffened gamma law gas.
For simplicity, we assume that the mixture of the two fluids, at the level
of a single grid block not meeting an interface, or for a cut grid block on one
side of an interface, is mixed at a molecular level. Thus we do not consider
turbulent modeling corrections to the grid level equation of state. The equation
of state for a mixture of stiffened polytropic gases is not a stiffened polytropic
gas, when the stiffening parameters p of the two species are unequal. As is
conventional, we impose pressure and temperature equilibrium for the grid cell
or cut cell mixture. The thermodynamic functions are then given as solutions
of an algebraic equation, and for the case to two fluids, as is considered here,
the algebraic equation is quadratic, and solvable in closed form.
The viscous stress tensor dij is expressed as
dij = d

vei
vej
+
xj
xi
16
2 vek
ij
3 xk
(2.18)
where d = k is the filtered dynamic viscosity. For simplicity, we take d

to be a global constant. A more fundamental theory of multifluid viscosity is
described in the book of Williams [56].
To derive the filtered energy equation (2.7), we write
1
(E + p)vi = evi + vk vk vi + pvi
2
(2.19)
1
= cv T vi + e vi + p vi + ( 1)cv T vi p vi + vk vk vi
2
(2.20)
1
e vei + (e
1)cev Tevei + p vei p vei ) + vek vek vei
= (cev Tevei + f
2
(2.21)
1
^
+ (c
ecev Tevei ) + (]
e vi ef
ei ) + (v^
ei ))
v T vi
v
k vk vi vek vek v
2
(2.22)
(H)
= (E + p)vei + (qi
1
(T )
+ qi ) kk vei + (]
e vi ef
ei ) .
v
2
(2.23)
We model the unclosed difference (]

e vi ef
ei ) as zero in (2.23).
v
2.2
Numerical Development for LES

The parabolic Navier-Stokes equations are solved via operator splitting,
with separate solution steps for the hyperbolic and pure diffusion parts of
the equations. The hyperbolic solutions are obtained by the front tracking
FronTier algorithm [10]. The interface hyperbolic updates are split into normal
and tangential operators defined at front points. The normal update uses
a predictor-corrector algorithm, with the predictor step a Riemann solution
17
using the left and right front states as input. The corrector step, based on finite
differences in characteristic coordinates, couples the normal direction front
propagation to signals coming from interior states. The interior hyperbolic
update uses a Godunov finite difference solver based on the MUSCL algorithm
[7, 57]. A sharp (tracked) interface in the hyperbolic update uses ghost cells
[27] in the interior state update to eliminate [38] transport related numerical
mass and thermal diffusion across the interface.
Physical transport for front states is introduced via finite differences in a
normal-tangential rectangular coordinate system at the front. Discretization
is outlined in Section 2.2.1, where the turbulent transport term is defined.
The parabolic update of the interior states, i.e. the inclusion of molecular
and turbulent transport terms for species diffusion, viscosity and thermal heat
conductivity, is via conventional finite differences. Irregular stencils for the
interior states, i.e. the stencils crossing the front, are treated conventionally,
without regard to the front or any ghost cell values. An explicit solver for
both the interior and the front state parabolic solvers, with possible time step
subcycling, is sufficient to allow a stable computation for most of the mesh
and transport parameter range considered.
For the highly heat diffusive plasma conditions, an implicit solver is used.
With transport terms strong enough to require an implicit solver, there is no
need for, or benefit from, tracking. Moreover, the formulation of the tracked
front states together with the interior states appears to be inconsistent with
common implicit solver algorithms. We allow selective untracking of specific
(primitive) variables within an overall tracked solution. In this algorithm, after
18
the interior and front state updates, the untracked front states (the primitive
variables that are not being tracked, for example energy, in the case p) are redetermined by one sided interpolation from neighboring interior states of the
same component. Here one sided means that the interpolating interior states
are restricted to have the same component as that of the front state in question.
In the unusual case of no nearby interior states with the required component, a
two sided interpolation is used, in which the component restriction is dropped.
As a test of this algorithm, we compared it to two sided interpolation, and
we compared it to an explicit algorithm (with tracking of the primitive energy
variable). For this case, extensive subcycling in the parabolic step for the
energy equation was needed for numerical stability. There was no notable
difference among these three algorithms. The FronTier numerical Schmidt
and Prandtl numbers are , and the code allows efficient simulation of any
desired (physical) Schmidt or Prandtl number.
We develop the subgrid scale dynamic model terms used in this thesis,
following and slightly extending ideas of [40, 42].
2.2.1
A Subgrid Scale Dynamic Model for the Momentum Equation
We use the trace-free Smagorinsky eddy viscosity model for the sub grid
scale (SGS) stress ij . The ij can be decomposed into an anisotropic (trace
19
zero) tensor (ija ) and an isotropic tensor (iji ), which are modeled separately:
ijM

ij
ij
e Seija + ij 2CI 2 |S|
e2 ,
= ij kk
+ kk
= ija + iji = 2CS 2 |S|
3
3
3
(2.24)
where Seija = Seij 3ij Sekk . and | | is defined by (2.4). The CS and CI are model
coefficients to be computed dynamically. In analogy to (2.18), we define

e
kt = CS 2 |S|
(2.25)
as the kinematic turbulent viscosity.

The key element of the dynamic model is the utilization of the data contained in the resolved field. This information is brought to bear by introducing
b than the resolved grid filter. We will
a test filter with a larger filter width
use a 2 2 mesh block average to define the test filter. Let a spatially testfiltered quantity be denoted by a caret. The test filtered stress Tij is defined
as:
[
Tij = v
i vj
and is modeled as:
ci v
cj
v
b
2ij b 2b be 2
be bea
a
i
b 2b
|
S|
S ij +
TM
=
T
+
T
=
2C
CI |S| .
S
ij
ij
ij
3
(2.26)
(2.27)
Using Germanos identity [14], the Leonard stress Lij can be expressed
20
as:
\
c
ci v
cj
vi vj
v
e
vi c
vej
[
Lij = Tij c
=
=
.
e
v
v
e
ij
i j
b
b
(2.28)
The right hand side is completely computable from the resolved variables. We
next introduce the ansatz that CS and CI are independent of the length scale.
In other words, the same CS and CI occur in (2.24) and (2.27). In this case,
where
be bea
\
e Sea 2CS
b 2b
Laij = Taij bija = 2CS 2 |S|
|
S|
S ij = CS Mija ,
ij
be bea
\
e Sea 2
b 2b
Mija = 22 |S|
|
S|
S ij .
ij
(2.29)
(2.30)
We would next solve (2.29) for CS . However, this equation corresponds to five
independent relations for CS and an algebraic solution is not possible.
We introduce an averaging operation h i. The specification of the average is problem dependent, as the universal definition of an ensemble average is
inconvenient to use. For the present problem, we regard this average as taken
over the symmetry variable , i.e., an average over circular arcs (constant radius). To assure numerical regularity, we also apply a convolution average in
the radial direction with a stencil extending 6r from current radius.
Applying this average to (2.29) and using least squares in the resulting
21
equations leads to the formula

P
h( Laij Mija )+ i
CS = P a a + .
h( Mij Mij ) i
(2.31)
Here the expression ( )+ means the positive part of the quantity in the
parenthesis. In other words, we clip negative values. This step is consistent
with methods proposed elsewhere [40]. The same method is used for the
coefficients of turbulent heat conduction and species concentration diffusion.
The turbulent viscosity determined by (2.31) and the related turbulent
transport coefficients determined below were monitored for the simulations of
the present study. The results were generally consistent with theoretical expectations. For example, the turbulent viscosity was respectively small (comparable, dominant) in relation to the molecular value of viscosity for the three
cases of DNS, LES near to DNS, and LES far from DNS.
To determine CI , we consider
be 2
[
i
i
e 2 + 2CI
b 2b
Likk = Tikk bkk
= 2CI 2 |S|
|S|
= CI Mkk
,
(2.32)
where a summation convention was used for the kk repeated indices and
be 2
[
i
e 2 + 2
b 2b
Mkk
= 22 |S|
|S|
(2.33)
This equation corresponds to one relation for CI , from which we obtain

hLikk i
CI =
.
i
hMkk
i
22
(2.34)
To discretize the elliptic operators associated with the molecular level

and subgrid scale transport terms, we use the same stencil at each front point
that was used in the front propagation. This stencil is rectilinear, and aligned
with the normal and tangential directions to the front at the front point in
question. We use three mesh points in each of the normal and the tangential
directions. Letting N and T denote derivatives in the normal and tangential
directions, we first write the discretization for N cN f for some function f and
spatially dependent coefficient c. The function is double valued at the front,
and so we specify, as an example, a discretization of the left value, fl . The
discrete operator has the form
c+1/2 (f+ f0l )c1/2 (f0l f ) = c+1/2 f+ (c+1/2 +c1/2 )f0l +c1/2 f . (2.35)
Here f0l and f0r are front states on the left and right side of the front, while
f and c1/2 are evaluated along the normal to the front at distances x or
1/2 x from the front.
The normal direction discretization for the right front state is similar,
and the tangential direction discretizations differ only in that all values are
associated with either the left or the right side of the front.
The discretization of the mixed partials is complicated by the use of a 5
point stencil and a fixed order of differentiation. We write
N cT f = T cN f + (N c)(T f ) (T c)(N f ) .
(2.36)
The normal sweep is performed in advance of the tangential sweep. Use of

23
(2.36) allows all mixed partial derivatives to be evaluated in the order of the
sweeps, first normal derivatives and then tangential derivatives. During the
normal sweep, we evaluate and store front values for N c and N f , which can
be differenced using obvious formulas during the tangential sweep to yield a
discrete expression for the LHS of (2.36).
2.2.2
A Subgrid Scale Dynamic Model for the Energy

Equation
We only consider the eddy diffusivity SGS model for the SGS heat trans(H)
port flux qi
with some modeling assumptions (e.g. negligible subgrid viscous

(T )
work and triple correlations). As in ref. [30], we set qi
(V )
= 0 = qi . The SGS
heat transport flux is modeled as:

(H)M
qi
= cep
= cep
e Te
CS 2 |S|
P rt xi
kt Te
,
P rt xi
(2.37)
(2.38)
where P rt is the SGS turbulent Prandtl number to be determined using a

dynamic model.
At the test filter level, the flux is defined as:
(H)
Qi
c v
ci
cp T
c
\
= c
,
p T vi
2
b
24
(2.39)
and is modeled as:

(H)M
Qi
= b
cb
ep
be be
b 2 |S|
CS
T
.
P rt xi
(2.40)
Again, using Germanos identity, we have
(H)
Li
(H)
Qi
d
(H)
qi =

cc
c v
cp T
ci
c\
c
c
cep Tee
vi
p T vi
\
e
= cep T vei
.
2
2
2
b
b
(2.41)
To determine P rt ,
(H)
Li
(H)M
= Qi
CS 2
\
(H)M
qi
=
P rt
\e!
be
CS (H)
be T
b 2b
e T
cep |S|
cb
ep |S|
=
M
,
xi
xi
P rt i
(2.42)
where
(H)
Mi
= 2
\e!
b
T
Te
2b b b
e
b
e
cep |S|
cep |S|
.
xi
xi
(2.43)
This equation corresponds to three independent relations and again a least

square approach is followed to calculate the model coefficient,
P
(H)
(H)
Mi Mi )+ i
P rt = CS P (H)
.
(H)
h( Li Mi )+ i
h(
Discretization is as in (2.35).
25
(2.44)
2.2.3
A Subgrid Scale Dynamic Model for the Concentration Equation
For SGS scalar transport, from a gradient transport modeling assumption,

()M
qi
e e
CS 2 |S|
kt e
=
,
Sct xi
Sct xi
(2.45)
where Sct is the SGS turbulent Schmidt number to be determined using a

dynamic model.
At the test filter level, the SGS scalar transport is defined as:
()
Qi
and is modeled as:

()M
Qi
c
ci
v
d
= vi
,
b
be be
b 2 |S|
CS
b
=
.
Sct xi
(2.46)
(2.47)
Using Germanos identity, we have
Li
()
Qi
d
()
qi =
\
c
c
ci
vi
v
vei e
de c
= e
vi
.
b
b
26
(2.48)
To determine Sct ,
()M
Li = Qi
CS 2
[
()M
qi
=
Sct
\e !
be
CS ()
be
e
b 2b
|S|
=
M
,
|S|
xi
xi
Sct i
(2.49)
where
()
Mi
= 2
\e !
b
be e
e
b 2b
|S|
|S|
.
xi
xi
(2.50)
This equation corresponds to three independent relations. Again, a least

square approach is followed to calculate the model coefficient,
Sct = CS
()
()
Mi Mi )+ i
.
P () ()
h( Li Mi )+ i
h(
27
(2.51)
Chapter 3
The Flow Instability Problem
The turbulent mixing considered here is initiated by a shock wave passing

through a layer separating two fluids of distinct densities. When the layer is
perturbed (or not normal relative to the shock wave), vorticity is deposited on
the interface by the shock passage. This vorticity causes the interface to roll up
and become unstable. Upon passage of a second shock wave, the interface enters an extremely chaotic regime. This is an example of a Richtmyer-Meshkov
(RM) instability.
We consider a circular geometry, with a converging circular shock at the
outer edge, and inside this, two fluids separated by a perturbed circular interface. The chaotic aspects of the mixing at a molecular level following reshock
challenge some conventional ideas of computational science while supporting
others. For this reason, the problem is of fundamental scientific interest, and
may shed light on differing views for the computation of turbulent mixing
flows.
28
3.1
Test Problem Setup
We consider a computational domain x, y [0, 25] [25, 25] (units of

cm). Time units are in micro seconds, and pressure is in Megabars. With r
denoting the radial coordinate in the x, y plane, the initial contact is perturbed
from a circle at r = 12.5. Outside this contact is a stiffened gamma law gas,
with parameters given by = 3.27, p = 0.1495, e = 1.25 1018 and
cv = 2.2 106, where the stiffened gamma law gas is defined by
e + e = cv T +
(3.1)
and
p + p = ( 1)cv T.
(3.2)
Here, is the adiabatic exponent and cv is the specific heat at constant volume
and e is the energy of formation and p has the dimensions of a pressure.
Inside the contact is also a stiffened gamma law gas, representing lucite, with
parameters = 1.85, p = 0.03036, e = 2.49 1018 and cv = 1.46 105 .
The time dependent boundary condition from the RAGE code [41] is located
at r = 24. The initial ambient pressure is p = 106 , and the initial pressure
at the boundary is p = 0.02648, giving rise to an inward propagating shock
at t = 0. The heavy fluid density is 7.282 and the light fluid density is 1.182,
giving an Atwood number A = 0.72. Initial and late time simulation density
plots of unregularized simulations (zero transport coefficients) are shown in
Fig. 3.1.
29
Figure 3.1: Initial (left) and late time (right) density plot for the Richtmyer-Meshkov fluid instability (unregularized, with zero physical transport coefficients) under study in this paper.
30
In the problem considered here, the flow 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 fluids, proceeds to the origin, reflects there and expands outward, recrossing the interface
region and finally exiting at the outer boundary. The interface between the
two fluids is defined as a 50% iso-concentration contour.
Due to the shock induced instability, the interface region expands into a
mixing zone, which has a very complex structure. Especially after the second
passage of the shock (the reshock or reflected shock passage), the mixing zone
becomes highly chaotic. The inner and outer edges of the mixing zone are
defined in terms of 5% and 95% volume fraction contours, after a spatial
average over the circular symmetry variable. The mixing zone is then defined
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 filter [11, 16, 59]. A space time plot of the shock trajectories and
mixing zone edges is shown in Fig. 3.2.
We have already observed [33] 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. 3.3, where the length / area ratio is shown in physical units
(left frame) and mesh units (right frame); transport coefficients have been set
to zero.
From Fig. 3.3, we observe that somewhat after reshock, the interface
length, in mesh units, occupies a constant ratio to the mesh area of the mixing
31
120
100
80
60
0
20
40
Time
10
15
20
Radius
Figure 3.2: Space time (r, t) contours of the primary waves, as detected by the
wave filter algorithm. These are the inward (direct) and outward (reflected)
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 ].
32
0.6
200 x 400
400 x 800
800 x 1600
1600 x 3200
15
grid length/grid area
interface length/mixing zone volume
20
10
20
40
60
time
80
100
200 x 400
400 x 800
800 x 1600
1600 x 3200
0.4
0.2
120
20
40
60
time
80
100
120
Figure 3.3: 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.
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 finest grid simulations. Further mesh refinement
studies will be needed to determine the evolution of mesh level complexity under continued mesh refinement. In any case, at the grid levels attained here,
the interface is mesh volume filling, cutoff by the mesh, and highly complex
or chaotic in nature.
3.2
Research Objectives
Our goal is accurate numerical solutions using feasible grids. For the
purpose, we verify the use of large eddy simulations (LES). We study mesh
refinement and mesh convergence and dependence on a range of Reynolds
numbers, including some within the regime of direct numerical simulation
33
(DNS), in which transport properties are fully resolved.

The fluid interface, at late time, is volume filling. The Reynolds number
and transport coefficients (viscosity, mass diffusion, and heat conductivity) are
given dimensionlessly as Re = UL/k , the Schmidt number Sc = k /D, and
the Prandtl number P r = k /. Here k is the kinematic viscosity, D the
kinematic mass diffusivity and =
cp
the kinematic thermal diffusion rate.
is the heat conductivity, the density and cp the specific heat at constant
pressure. U and L are characteristic velocity and length scales. Gases typically
have Schmidt and Prandtl numbers of the order of unity, while liquids typically
have Schmidt numbers in the range Sc 4 102 to 104 and Prandtl numbers
in the range P r 10 to 102 [1, 48]. Dense plasma transport coefficients have
been estimated as Sc 0.6 to 1.5 [52] for a range of plasma conditions relevant
to NIF experiments. Plasma Prandtl numbers are dependent on ionization
levels, nuclear charge, and temperature. A representative plasma value P r
104 can be inferred from the viscosity and heat conductivity values quoted by
Drake [9]. The very high level of thermal conductivity reflected in this value
results from the transport of free electrons, and a time scale sufficient for
the thermalization that allows a single temperature description of the plasma.
We make no comments on this time scale. Numerical Schmidt and Prandtl
numbers are generally not documented. If we assume a numerical shock width
of 1.5x and a numerical contact width of 5x for untracked simulations,
numerical transport coefficients in the range of 0.3 would be reasonable. On
this basis, we consider the transport cases l, g, p from Table 3.1. We allow
physical parameters (Re, Sc, P r) to vary by three orders of magnitude, and
34
Table 3.1: Transport coefficients considered in this paper, arranged in order of

increasing diffusivity. The presumed typical numerical transport parameters
are shown for reference only and are not explored here.
case
l (liquid)
g (gas)
n (numerical)
p (plasma)
Schmidt
103
1
0.3
1
Prandtl
50
1
0.3
104
explore mesh refinement up to 3200 zones per linear dimension. To keep the
computational burden manageable, we restrict the simulations to 2D and we
restrict the 2D mesh resolution to the finest resolution presently used for 3D
simulations.
The emphasis on the joint probability distribution function (pdf) of concentration and temperature as micro scale observables is motivated by problems in turbulent combustion [47], where these variables affect the local flame
speed and the overall flow.
The subgrid models are not original here, but their use in flow simulations
having numerically sharp gradients appears to be new. In this sense, we are
outside of the known domain of validity of the subgrid models, and the present
work serves as a verification (i.e. mesh convergence with full resolution of the
momentum equation) study.
At least for the present problem, with its somewhat modest number of
initial modes, the joint pdf for concentration and temperature are subject to
statistical fluctuations. In other words, the spatial averaging over the mixing
zone is not sufficient to obtain statistical convergence, and an ensemble of
35
simulations would be needed to obtain statistically converged mixing statistics.

This can be observed from Fig. 3.1, where we see a significant variation in
the size of the coherent mixing structures, and a relatively small sample of
the larger sized ones, whose size appears to reflect the t = 0 perturbation
wave length. For this reason, we wish to introduce some degree of averaging
into our analysis of the pdfs. Looking to the various applications in which
turbulent mixing plays a role, chemical reactions stand out. We are involved in
combustion modeling for scram jet design, type Ia supernova studies, and ICF
motivated studies. Accordingly, we consider the chemical reaction rate w of a
hypothetical reaction A + B C with a hypothetical activation temperature
TAC to assess convergence of the pdfs. The coefficient of variation of w, in its
dependence on space at fixed time within the mixing zone, is generally over
100%. Thus spatial averaging is needed.
With this definition of spatially averaged convergence, the statistical fluctuations are reduced but not eliminated. We present (as error bars in plots)
the fluctuations (2) associated with variation across an ensemble of sparially averaged quantities. We adopt a pdf for the (single realization) reaction
R
rate values w, with a probability measure dP (w). Then 1 = dP (w) and the
R
spatial average of w is hwi = wdP (w). We introduce the notation dP x (w),
x = c, m, f for the w-pdf for the coarse, medium and fine grids. Then we
R
define the relative error as w|dP c(w) dP f (w)|/hw f i for the coarse to fine
error, and similarly for the medium to fine error. These continuum expres-
sions are interpreted in terms of binning the data, generally into 10 bins per
variable. We observe mesh convergence. Other definitions of convergence give
36
similar results, except that when the error is well below the level of ensemble
variability (realization to realization variation), convergence may be obscured
by chance cancellations giving unrealistically small net error on coarse grids.
The error definition chosen minimizes such cancellations.
To summarize, the main goal of this paper is to introduce and verify parameterized subgrid models for turbulent mass, momentum and thermal diffusion which will capture unresolved diffusive phenomena as it impacts coarse
grid scales in a LES having steep numerical gradients. In this way, we plan to
achieve LES which are converging relative to both the macro and the micro
observables mentioned above. The simulations are more efficient than those
of conventional turbulence models in the narrow width in mesh units that
they allow for sharp gradient concentration transitions. They model turbulent transport with microscopic observables correctly computed, in contrast
to many capturing simulations. These subgrid models are applied here in a
new context, namely to a front tracking and shock capturing Godunov scheme
which maintains sharp gradients.
The front tracking code FronTier can achieve arbitrarily high Schmidt
and Prandtl numbers numerically without a requirement for mesh refinement
beyond that needed to resolve the momentum equation. This is not the case
for most (untracked) simulation codes. Numerical mass and thermal diffusion
arises primarily within the Eulerian hyperbolic step, due to solution averaging over grid blocks associated with transport by a nonzero velocity field (i.e.
hyperbolic transport). This apparently universal feature of untracked Eulerian conservative capturing codes is circumvented with front tracking. Subgrid
37
models can only add diffusion, and can never remove it. In this sense, a subgrid model for mass or thermal diffusion cannot cure the problem of excess
numerical diffusion to achieve accurate modeling of atomic scale mixing. A
conventional untracked code must be run in an over resolved mode, with
more resolution than the momentum equation requires, to achieve low levels
of numerical mass or thermal diffusion in the concentration and energy equations. In other words, conventional untracked codes require that the simulations be resolved or convergent in terms of the continuity, species and energy
equations as well as for the momentum equation. The use of front tracking,
however, avoids or reduces numerical diffusion in the species, continuity, and
energy equations, even for coarse grids, and thus requires DNS or LES type
convergence of the momentum equation alone.
We also point to the conclusions of Chapter. 4, in which the macro variables (mixing zone edge positions, shock trajectories, etc.) are insensitive to
physical or numerical modeling issues, and thus presumably to the choice of
the subgrid model for viscosity, or to the viscosity and Reynolds number itself.
Thus our convergence studies will focus primarily on the micro variables: the
joint pdfs for species concentration and temperature and on a typical chemical
reaction rate. We will examine the degree to which these are insensitive to
Reynolds number for LES, and to mesh for a fixed Reynolds number LES.
There is a large literature concerning turbulent mixing in RM unstable
flows. Most of this literature focuses on macro observables, such as the mixing
zone edges and shock trajectories. For RM mixing, these macro observables are
insensitive, in that we find generically agreement among theory, experiment
38
and numerics, cf. [31].

The micro observables (e.g. joint temperature and species PDFs) have
received relatively less attention, but we can cite a previous study [30], with
conclusions distinct from ours, for a different RM flow. Among experimental
studies of micro observables for RM instabilities, we mention two studies of
related but distinct flow problems [32, 55]. These experimental studies appear
to be qualitatively consistent with our conclusions regarding the mixing, when
compared at comparable times and Schmidt numbers. Quoting [32], Schmidt
number plays a role in turbulent mixing of high-Reynolds flows.
The critical dependence of molecular levels of mixing on physical transport (and for under resolved capturing simulations, on numerical transport)
is illustrated in Fig. 3.4. This dependence is one of the central points of this
thesis.
39
Figure 3.4: Chemical reaction rate at t = 90 for three representative values for
the transport parameters, and each for three values of the Reynolds number.
The activation temperature is chosen as 15, 000oK. The point of this plot is
the strikingly different values obtained for hwi for the nine physical regimes
considered here.
40
Chapter 4
Mesh Convergence for Macro Observables
We define a (time dependent) length scale L to be the width of the mixing

p
zone, and the velocity scale U to be the turbulent velocity U = hv 2 i. The
angle brackets h i denote a spatial avergae over the mixing zone. We observe
below that ensemble averages are also relevant. We define v = v hvi. We

also define Remesh = Ux/k .
The objective of the present chapter is to compute the large scale solution
features accurately. This includes the trajectories of the principal waves, as
illustrated in Fig. 3.2. This objective is related to a systematic convergence
study [34, 59]. In that study, we found statistical convergence for many mean
flow variables, which define what we call the macroscopic description of the
flow. In Fig. 4.1, we plot the time integrated relative wave discrepancy defined in terms of the mixing zone edge positions for a variety of mesh levels
and for transport coefficients from Table 3.1. Similar convergence properties
have been obtained for other macro variable solution errors such as the mean
densities and velocities for each phase. The error (or discrepancy) is deter41
mined by comparison of the simulation to a fine grid (3200 1600) simulation

having zero transport coefficients. The reported discrepancy is thus a mixture
of mesh errors and discrepancies associated with modification of the transport
coefficients from a nominal value (zero). From Fig. 4.1, and related studies
[34, 41, 59] we conclude that the macro observables are insensitive to both numerical and physical parameters, except for case p with low Re (large transport
parameters).
relative wave discrepancy
0.4
400x800 Liquid case
400x800 Gas case
400x800 Plasma case
800x1600 Liquid case
800x1600 Gas case
800x1600 Plasma case
1600x3200 Gas case
0.3
0.2
0.1
10
-2
10
-1
Kmesh
10
10
Figure 4.1: Time integrated relative discrepancy in the mixing zone edge locations as compared to a fine grid, zero transport simulation. The discrepancy,
for cases l and g, is mainly due to mesh errors and is partly due to modification of physical parameters. The discrepancy decreases with mesh refinement,
uniformly as physical parameters are varied. The plasma case p shows convergence, but to a Reynolds number dependent limit distinct (for small Re) from
the zero transport case.
42
16
10
8
400x800 Liquid case

400x800 Gas case
400x800 Plasma case
800x1600 Gas case
1600x3200 Gas case
0.4
0.3
12
Cmesh
0.5
400x800 Liquid case

400x800 Gas case
400x800 Plasma case
800x1600 Gas case
1600x3200 Gas case
14
0.2
6
4
0.1
2
0
10-2
10-1
Kmesh
100
101
0 -4
10
10
-3
10
-2
10
-1
Figure 4.2: Left: plot of C /x vs. Kmesh for a range of mesh levels and
transport parameters cases l, g, p from Table 3.1, at t = 90, r = rmid . Right:
the same data replotted as C vs. K .
4.1
Correlation and Diffusion Length Scales

We introduce a correlation length scale C to characterize the microstruc-
ture of mix. The correlation length is defined in terms of the probability of

exit distance from a given phase or mean distance to the complementary
phase, introduced [49, 50] for models of opacity, and studied [17, 34, 35] as a
measure of fine scale mixing length. For random points situated on a radius
r within the mixing zone, the exit probability data is collected into bins, each
holding the data for an interval of possible exit lengths. Using the probability
R
measure d, as defined by the binned data, we define C = d. See Fig. 4.2.
We assess interface convergence in terms of the behavior of C . The
regime of a resolved momentum equation occurs to the right side of each

frame of Fig. 4.2. In this regime, the right frame (with scales independent
of x) suggests mesh convergence to a limit, with weak dependence on Sc
and P r. In the left frame of Fig. 4.2, we scale out the mesh dependence, and
43
l
Re
300
6000
600K
D/Dt
0.031
0.002
1.54 105
g
/t
0.051
0.003
3.41 105
D/Dt
39.2
1.82
0.015
p
/t D/Dt
10.4
732
0.307 9.57
0.002 0.040
/t
4.16 106
1.32 104
27.1
Table 4.1: Ratios of dimensionless molecular to turbulent Schmidt and Prandtl

numbers for various cases, all with the finest grid. Ratios large relative to
unity indicate resolution of the associated transport equation. The momentum
equation has full resolution for Re 300 only.
observe that C
mesh
is only weakly dependent on x in the LES regime (left
part of the frame). In other words, C C
mesh x,
for an LES simulation,
is determined mainly by the mesh level.

We introduce the time dependent molecular diffusion length scales D =
2(D(t t0 ))1/2 , and T = 2((t t0 ))1/2 ( = /cp ), where t0 is the time of
reshock. The corresponding turbulent length scales are Dt = 2(Dt (t t0 ))1/2 ,
and Tt = 2(t (t t0 ))1/2 . Here Dt = kt /Sct and t = kt /P rt are kinematic
turbulent transport coefficients (species and temperature). Sct and P rt are
turbulent transport coefficients and kt is the kinematic turbulent viscosity,
all defined in Section 2.2. The ratios (D + Dt )/C and (T + Tt )/C are
dimensionless measures of the concentration and thermal mixing levels due to
combined effects of molecular and turbulent transport. We plot these ratios
vs. K in Fig. 4.3 for a variety of meshes and for transport coefficients from
Table 3.1. The large K asymptote (right side of each frame of Fig. 4.3)
represents the limit of a resolved momentum equation. This limit is converged
to a grid independent value, which depends on Sc and P r.
44
8
7
(D+Dt)/C
6
5
4
(T+Tt)/C
400x800 Liquid case

400x800 Gas case
400x800 Plasma case
800x1600 Gas case
1600x3200 Gas case
3
2
10
10
10
10
10
400x800 Liquid case

400x800 Gas case
400x800 Plasma case
800x1600 Gas case
1600x3200 Gas case
1
0 -4
10
10
-3
10
-2
10
10 -4
10
-1
10
-3
10
-2
10
-1
Figure 4.3: The ratios (D + Dt )/C and (T + Tt )/C vs. K for several
mesh levels and for transport coefficients cases l, g, p from Table 3.1. The
values to the far right in each frame are momemtum-resolved, and appear to
be mesh converged for the finer meshes.
The ratios are also mesh convergent for higher Re (in the LES regime)
and nearly independent of Re. Fig. 4.3 shows the success of the subgrid model,
with the mesh converged LES in excellent agreement with momentum-resolved
simulations.
It is interesting to compare the ratios of molecular to turbulent length
scales. The factor t t0 cancels and the ratio is just the ratio of the diffusion
coefficients, molecular to turbulent. In Table 4.1, we give the ratios of molecular to turbulent transport properties. The ratios are mesh and Reynolds
number dependent. Turbulent transport terms are defined in terms of subgrid
velocity correlations (with concentration, velocity, or temperature) and are
thus logically independent of molecular transport. We tabulate results for the
finest mesh only. For the momentum-resolved simulations, the turbulent viscosity terms are neglible, and for high Reynolds number, the turbulent terms
45
are dominant. Consistent with this picture, we note the virtual identity of the
high Reynolds data in Fig. 4.3, at least for the liquid and gas cases.
46
Chapter 5
Convergence of Micro Observables
5.1
Joint PDFs
The joint pdf for the temperature and species mass concentrations of the
fluid mixture is defined as a function of time, assuming that the probability data is collected from the spatial variation of a single realization (single
realization of the intial conditions) within the mixing zone. Mixed cells are
not averaged, but each cell fraction contributes its own concentration fraction and temperature with its own probabilities (proportional to area). The
concentration fractions and temperatures are then binned, with 10 bins per
variable.
The liquid and gas joint pdfs are bimodal. Both are concentrated near a
curve in concentration-temperature space, joining the light to the heavy fluid
concentrations. The origin of this shape can be explained by shock heating,
which tends to heat the heavy fluid more strongly, according to the properties of the assumed EOS for the two fluids. This trend is reinforced by the
initial conditions; initially the heavy fluid is hotter. Then portions of the
47
1600x3200, Re=6.2k, Sc=1000, Pr=50
0.4
probability
0.3
0.2
0.1
0
9577
1
temperature
24705
Light fluid mixture fraction
Figure 5.1: Case l. The joint pdf of light species concentration and temperature
at time t = 90. The data has been collected into 10 10 bins. The mesh is
1600 3200 and Re 6000.
heavy and light fluid diffuse into one another, so that the temperature pdf at
fixed concentration is determined from the temperature pdf of the pure fluids before mixture through diffusion. The plasma joint pdf is bimodal in its
dependence on concentrations, while its temperature dependence is unimodal.
See Figs. 5.1, 5.2, 5.3. The data presented is from the time t = 90, which is
the beginning of the chaotic stage of interface development.
The joint pdfs for temperature-concentration with common physics at
three mesh levels are qualitatively similar, and show signs of convergence. See
Table 5.1. The mixed region displays some number of large blobs of poorly
mixed light material. The relative size of these regions is statistically variable.
The errors quoted are measured in the Kolmogorov-Smirnov metric, i.e., the
L norm of the difference of the associated probability distribution functions.
48
1600x3200, Re=6.8k, Sc=1, Pr=1
0.4
probability
0.3
0.2
0.1
0
10405
1
temperature
21543
Figure 5.2: Case g. The joint pdf of light species concentration and temperature at time t = 90. The data has been collected into 10 10 bins. The mesh
is 1600 3200 and Re 6000.
1600x3200, Re=3.4k, Sc=1, Pr=0.0001
0.4
probability
0.3
0.2
0.1
0
15375
1
temperature
16511
Figure 5.3: Case p. The joint pdf of light species concentration and temperature at time t = 90. The data has been collected into 10 10 bins. The mesh
is 1600 3200 and Re 3000.
49
l
Re
300
6000
600K
c to f m to f c to f m to f c to f m to f
0.24
0.13
0.08
0.06
0.57
0.31
0.07
0.06
0.09
0.05
0.68
0.41
0.26
0.06
0.22
0.09
0.14
0.10
Table 5.1: Mesh errors for the joint temperature-concentration pdfs to illustrate possibilities of mesh convergence. Comparison is coarse (c) to fine (f)
and medium (m) to fine with pdfs compared using the Kolmogorov-Smirnov
metric. Various physical cases reported.
We compute with three distinct meshes (400800, 8001600 and 16003200)
for the cases l, g, and p. Thus there are two levels of error for each case.
The listed Reynolds numbers are approximate, and the refinements are
performed (within each series of approximately constant Reynolds number)
with identical molecular level transport coefficients. Thus the mesh refinement
occurs within identically specified physics. The observed Reynolds numbers
and the subgrid scale turbulent transport parameters are mesh dependent, and
do vary within this comparison.
5.2
Concentration Moments
Although the errors in the pdfs decrease with mesh refinement, they are
not uniformly small, and in the worst cases are as large as the pdfs themselves.
For this reason, we next consider a more highly averaged analysis of the same
simulation data.
50
The mean molecular mixing fraction , defined [58] as
hf1 f2 i
,
hf1 i hf2 i
(5.1)
is a common measure of mixing. Here fk is the mass concentration fraction for

the species k. Perhaps the best way to understand is through its role in the
specific chemical reaction production rate w for a simple reaction A + B C,
namely [3]
w = const. f1 f2 exp(TAC /T ) ,
(5.2)
where TAC is an activation temperature and w is set to zero if T < TAC .

The constant in (5.2) is dimensionless. In this formula, we note that occurs
naturally as a factor in the mean value hwi for w. We use moments () to
analyze the variability of the concentration, but due to the strongly nonlinear
dependence of w on T , we prefer to model T using its pdf. In this way, we
analyze the pdf for w with no use of a closure model.
In order to compute the mean value of w, we need to perform the statistical average defining at fixed T . This means that we bin together data
points with T in a common range (bin), and in this restricted ensemble, we
compute means to define . Let hiT denote the expectation in the fixed T
spatial ensemble. It is defined as the sum of its argument over all sample
points in a fixed T bin divided by the number of sample points in that bin.
The result is (T ), as a function of temperature T , where
(T ) =
hf1 f2 iT
.
hf1 iT hf2 iT
51
(5.3)
l
300
6000
600K
0.09 0.09 0.03 0.01 0.23

0.07 0.05 0.14 0.08 0.36
0.11 0.06 0.12 0.05 0.07
0.06
0.15
0.07
Table 5.2: Mesh errors for (T ), defined by the weighted L1 norm (5.4). Comparison is coarse (c) to fine (f) and medium (m) to fine. Various physical cases
reported.
The result is shown in the nine frames of Fig. 5.4. We note that the gas
case has values (T ) 0.6 0.8 over most of the temperature range. Aside
from a peak (T ) value associated with the lowest and highest T values (caused
by unmixed fluid), most of the (T ) range displays very little temperature
dependence. Because the defined conventionally by a spatial average hi
(not constrained to a fixed T value) as in (5.1) is distinct from (T ), we
include the value in each frame of Fig. 5.4. The values are generally
lower than those obtained in other simulations for related problems. This fact
reflects the elimination of numerical mass diffusion in our simulations and a
greater occurrence of blobs of unmixed or poorly mixed fluid than is commonly
observed numerically.
We supplement the visual convergence shown with a quantitative estimate

of the error in (T ). See Table 5.2. Here the error is a weighted L1 norm,
e=
|c (T ) f (T )|(dP c(T ) + dP f (T ))/2
9
X
wi |ic if |
(5.4)
i=0
where wi is the weight of the corresponding temperature bin, calculated (for

52
l
Re
300
6000
600K
0.29
0.16
0.06
0.03
0.07
0.01
0.08
0.07
0.10
0.06
0.20
0.08
0.13
0.03
0.18
0.07
0.03
0.04
Table 5.3: Mesh errors for the specific chemical production w for an activation
temperature TAC = 8, 000oK. The comparison, based on the Kolmogorov-Smirnov matric, shows mesh convergence. The comparison is coarse mesh (c)
to fine (f) and medium (m) to fine. Various physical cases reported.
the coarse to fine comparison) as
wi =
5.3
uci + ufi
.
2
(5.5)
Temperature PDF
Encouraged by the fact that the variable (T ) is mesh convergent, as a
factor contributing to the chemistry production term w, we next take the step
of analyzing the pdf for T . By definition, this pdf (which we denote p(T )), is
the number of sample points in a T bin per degree size for the width of the
bin in degrees Kelvin. We plot the pdf for T , in Fig. 5.5, for various physical
cases.
Again the convergence properties are visually encouraging. Pursuing
this point, we tabulate the mesh convergence errors, as measured in the
Kolmogorov-Smirnov metric, for the pure T pdf, to show grid convergence,
see Table 5.4. The errors for the T pdf are defined as in (5.4).
53
0.4
400x800 =0.25
800x1600 =0.16
1600x3200 =0.16
14000
16000
18000
temperature (K)
0.4
20000
12000
14000
16000
18000
temperature (K)
0.6
0.4
20000
0.8
0.8
0.8
0.6
0.6
0.6
400x800 =0.61
800x1600 =0.58
1600x3200 =0.56
0.2
0
13500
15000
16500
temperature (K)
18000
()
0.4
0.4
400x800 =0.48
800x1600 =0.48
1600x3200 =0.50
0.2
0
19500
12000
14000
16000
18000
temperature (K)
0.8
0.8
0.8
0.6
0.6
0.6
400x800 =0.43
800x1600 =0.39
1600x3200 =0.38
10900
10950
11000
temperature (K)
11050
()
()
0.2
400x800 =0.30
800x1600 =0.36
1600x3200 =0.42
0.2
0
15600
15800
16000
16200
temperature (K)
16400
15000
18000
24000
400x800 =0.45
800x1600 =0.47
1600x3200 =0.52
12000
15000
18000
temperature (K)
21000
0.4
0.2
0
10000
400x800 =0.53
800x1600 =0.60
1600x3200 =0.62
12500
15000
17500
temperature (K)
Figure 5.4: Plot of (T ) vs. T . Cases l, g, p (rows 1 to 3), with Re increasing, left to right. Error bars indicate coarse grid ensemble fluctuation (2)
centered at the ensemble mean, for cases l, g only.
54
21000
temperature (K)
0.2
20000
0.4
12000
0.4
0.4
400x800 =0.45
800x1600 =0.48
1600x3200 =0.50
0.2
400x800 =0.42
800x1600 =0.42
1600x3200 =0.35
0.2
()
()
12000
0.6
()
0.6
0.2
()
0.8
0.8
()
()
0.8
20000
0.5
0.4
0.3
0.2
0.1
12000
14000
16000
18000
temperature (K)
0.1
0.3
0.2
0.1
13500
15000
16500
temperature (K)
18000
0.4
0.3
0.2
0.1
12000
14000
16000
18000
temperature (K)
0.2
0.1
10900
10950
11000
temperature (K)
11050
Probability Density (/100K)
0.3
0.2
0.1
15000
18000
21000
temperature (K)
24000
400x800
800x1600
1600x3200
0.4
0.3
0.2
0.1
12000
15000
18000
temperature (K)
21000
0.6
400x800
800x1600
1600x3200
0.5
0.4
0.3
0.2
0.1
0
12000
0.5
20000
0.6
0.4
0.3
0.6
400x800
800x1600
1600x3200
0.5
19500
400x800
800x1600
1600x3200
0.5
0.4
12000 14000 16000 18000 20000
0.4
0.2
400x800
800x1600
1600x3200
0.5
temperature (K)
0.6
0.3
0.6
400x800
800x1600
1600x3200
0.5
0.4
20000
0.6
0.5
15600
15800
16000
16200
temperature (K)
16400
0.6
400x800
800x1600
1600x3200
0.6
400x800
800x1600
1600x3200
0.6
400x800
800x1600
1600x3200
0.5
0.4
0.3
0.2
0.1
0
10000
12500
15000
Figure 5.5: Plot of T pdf. Cases l, g, p (rows 1 to 3), with Re increasing, left
to right.
55
17500
temperature (K)
20000
l
Re
300
6000
600K
0.13
0.08
0.05
0.05
0.57
0.31
0.07
0.03
0.07
0.04
0.68
0.41
0.27
0.06
0.23
0.07
0.13
0.09
Table 5.4: Errors for the temperature pdfs measured in the Kolmogorov-Smirnov metric show mesh convergence. Comparison is coarse mesh (c) to fine
(f) and medium (m) to fine. Various physical cases reported.
We note an important distinction between the case p and the cases l and
g. For the high thermally diffusive case p, the range of temperatures spatially
is very narrow. Accordingly, the convergence of the pdfs in this case is rather
like convergence of near delta functions, and as the mesh is varied, there can
even be only partial overlap in the temperature ranges observed. The Reynolds
number dependence is significant in case p but is weak in cases l and g.
5.4
Chemical Production
Finally, we combine the separate analysis of (T ) and T to estimate a
convergence rate for the pdf for the specific chemical production w, assuming
arbitrarily activation temperatures TAC = 8, 000o K and 15, 000o K. Here,
TAC = 8, 000o K is lower than the range of observed temperatures, while
TAC = 15, 000oK is in the middle of the observed temperatures (cases l, g
only).
We list convergence properties for the w pdf, based on the KolmogorovSmirnov metric in Tables 5.3 and 5.5. The evidence is encouraging, and sug56
Re
300
6000
600K
l
g
2/ hwi c to f m to f 2/ hwi c to f
0.24
0.04
0.03
0.29
0.03
0.49
0.04
0.04
0.45
0.03
0.25
0.09
0.03
0.25
0.09
m to f
0.02
0.04
0.05
Table 5.5: Relative coarse grid ensemble fluctuations (2) divided by ensemble mean, and mesh errors for the specific chemical production w for an activation temperature TAC = 15, 000oK. Comparison based on the Kolmogorov-Smirnov metric. Comparison is coarse mesh (c) to fine (f) and medium (m) to
fine. Only cases l and g are reported.
gests convergence, subject to additional mesh refinement. The random variable
w displays not only spatical variablity, but also ensemble (realization to realization) variability. Because the temperature range in case p is small, and
varies (even to the point of being non-overlapping) as Re and the mesh is
varied, it does not seem realistic to choose a TAC within the range of observed
temperatures. Accordingly, we omit case p from Table 5.5.
5.5
Statistical Fluctuation vs. Mesh Convergence

Close examination of Tables 5.4 - 5.5 suggests anomalies with the conver-
gence, with a few cases showing only marginal convergence. For this reason
we examine the mesh convergence more closely. First, we looked at errors
or randomness associated with the binning process. The results did not show
significant sensitivity to the binning. Next we examined the radial dependence
of the convergence. We found a significant radial dependence on the temperature, a consequence of the fact that the heavy material is hotter than the light.
57
0.04
<w>
0.03
0.02
0.01
radius
Figure 5.6: Radial dependence for hwi for the case l, Re 6000.
So the temperature and w, considered as a function of r, is biased towards the
outer portion of the mixing zone. See Fig. 5.6. Finally, we considered the relative importance of mesh errors and statistical fluctuations. For the coarsest
grid, and for cases l and g only, we introduced a modest ensemble (6 realizations) defined by randomly selected initial perturbations of the interface.
We find that the mesh errors are very small, and are dominated by random
fluctuations (magnitude defined as 2 for this small ensemble). Thus we believe that the observed error fluctuations are probably chance events, rather
than indicating convergence or its absence. The spatial variabiity of the point
values of w, as indicated by its spatial coefficient of variation, typically 100%
or more, shows the importance but not the sufficiency of spatial averages.
58
Chapter 6
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 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 DiPerna [45, 46] and Ding and co-workers [4, 8],
the first and easier step is to show existence of the generalized solution. The
more difficult 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 compact59
ness 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 scientifically 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 scientific
truth or not.
60
Chapter 7
Conclusion
Our most important conclusion is a converging trend for the micro observables for a strongly chaotic mixing flow: reshocked Richtmyer-Meshkov
instability. The micro variable mesh convergence is partly obscured by statistical fluctuations, as noted below and in the body of the paper. New computational strategies and new methods of data analysis were required to reach
this goal. We have combined (and enhanced) the resolution of sharp gradients
near discontinuities of capturing codes with the subgrid physics modeling accuracy of turbulence codes. The convergence properties are basically uniform
in Schmidt and Prandtl number. As far as the authors are aware, results of
this nature have not been reported by others.
A striking feature of the results is their relative independence on the
Reynolds number, allowed to vary by a factor of 2000. Also significant is the
strong dependence of the mixing properties of the simulation on the physical regimes with their distinct Schmidt and Prandtl numbers. The impact of
diffusion within the simulations is generally lower than commonly reported
61
on the basis of under resolved untracked capturing codes. Such codes substitute numerical transport properties for physical ones. The actual numerical
transport values are not documented, but it is safe to guess that the effective
numerical Schmidt and Prandtl numbers are mainly less than one, and thus
more diffusive than the present gas case.
These trends in the diffusive properties of the solutions are predicted by
the trends in the (D + Dt )/C and (T + Tt )/C ratios. In other words,
elementary one dimensional diffusion analysis, coupled with a statistical characterization of the flow geometry, is sufficient to predict the major trends
observed. We refer to the plot of these ratios in Figs. 4.3.
As new modes of data analysis, we have introduced the importance of
convergence for the microscopic variables. These are subject to statistical
fluctuations which obscure mesh convergence. Some level of averaging is desirable, but we preserve important nonlinear functionals of the data, for which
simple averaging is meaningless. This leads to the ensemble labeled by concentration and temperature values rather than by spatial coordinates. Using this
framework and the typical observable of a chemical reaction rate, we observe
a reduction of statistical fluctuation and an appearantly converged pdf for the
specific chemical reaction rate.
62
Bibliography
[1] K. Asano. Mass Transfer: From Fundamentals to Modern Industrial Applications. Wiley-VCH, 2006.
[2] G. K. Batchelor. Small-scale variation of convected quantities like temperature in turbulent fluid. J. Fluid Mech, 5:2961, 1959.
[3] R. S. Cany and K. Mastorakos. An Introduction to Turbulent Reacting
Flows. Imperial College Press, London, 2008.
[4] G.-Q. Chen. Convergence of the Lax-Friedrichs scheme for isentropic gas
dynamics III. Acta Mathematica Scienitia, 6:75120, 1986.
[5] I-L. Chern and P. Colella. A conservative front tracking method for hyperbolic conservation laws. LLNL Rep. No. UCRL-97200, Lawrence Livermore National Laboratory, 1987.
[6] I-L. Chern, J. Glimm, O. McBryan, B. Plohr, and S. Yaniv. Front tracking
for gas dynamics. J. Comput. Phys., 62:83110, 1986.
[7] P. Colella. A direct Eulerian MUSCL scheme for gas dynamics. SIAM
journal on scientific computing, 6(1):104117, 1985.
[8] 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.
[9] Paul Drake. Hign energy Density Physics: Fundamentals, Inertial Fusion,
and Experimental Astrophysics (Shock Wave and High Pressure Phenomena). Springer Verlag, 2006.
[10] Jian Du, Brian Fix, James Glimm, Xicheng Jia, Xiaolin Li, Yunhua Li,
and Lingling Wu. A simple package for front tracking. J. Comput. Phys.,
213:613628, 2006. Stony Brook University preprint SUNYSB-AMS-0502.
63
[11] 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-AMS04-09 and LANL report No. LA-UR-04-2989.
[12] R. P. Fedkiw, T. Aslam, B. Merriman, and S. Osher. A non-oscillatory
Eulerian approach to interfaces in multimaterial flows (the ghost fluid
method). J. Comput. Phys., 152:457492, 1999.
[13] Ronald P. Fedkiw and Xu-Dong Liu. The ghost fluid method for viscous
flows. In Progress in Numerical Solutions of Partial Differential Equations, France. Arachon, 1998.
[14] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamic subgrid
scale eddy viscosity model. Phys Fluids A, 3:17601765, 1991.
[15] J. Glimm, M. J. Graham, J. W. Grove, X.-L. Li, T. M. Smith, D. Tan,
F. Tangerman, and Q. Zhang. Front tracking in two and three dimensions.
Comput. Math. Appl., 35(7):111, 1998.
[16] 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, 26:666
697, 2004. University at Stony Brook Preprint Number SB-AMS-03-11,
Los Alamos National Laboratory number LA-UR-03-2921.
[17] J. Glimm, J. W. Grove, X. L. Li, W. Oh, and D. H. Sharp. A critical
analysis of Rayleigh-Taylor growth rates. J. Comput. Phys., 169:652677,
2001.
[18] J. Glimm, J. W. Grove, X.-L. Li, K.-M. Shyue, Q. Zhang, and Y. Zeng.
Three dimensional front tracking. SIAM J. Sci. Comp., 19:703727, 1998.
[19] J. Glimm, J. W. Grove, X.-L. Li, and D. C. Tan. Robust computational
algorithms for dynamic interface tracking in three dimensions. SIAM J.
Sci. Comp., 21:22402256, 2000.
[20] J. Glimm, J. W. Grove, X.-L. Li, and N. Zhao. Simple front tracking.
In G.-Q. Chen and E. DiBenedetto, editors, Contemporary Mathematics,
volume 238, pages 133149. Amer. Math. Soc., Providence, RI, 1999.
64
[21] J. Glimm, E. Isaacson, D. Marchesin, and O. McBryan. Front tracking

for hyperbolic systems. Adv. Appl. Math., 2:91119, 1981.
[22] J. Glimm, X.-L. Li, and Y.-J. Liu. Conservative front tracking in higher
space dimensions. Transactions of Nanjing University of Aeronautics and
Astronautics, 18:115, 2001. Proceedings of International Workshop on
Computational Methods for Continuum Physics and Their Applications
(IWCCPA), Nanjing, China.
[23] J. Glimm, X.-L. Li, and Y.-J. Liu. Conservative front tracking in one
space dimension. Contemporary Mathematics, 295:253264, 2002. Proceedings of the Joint Summer Research Conference on Fluid Flow and
Transport in Porous Media: Mathematical and Numerical Treatment.
Report SUNYSB-AMS-01-16.
[24] J. Glimm, X.-L. Li, Y.-J. Liu, Z. L. Xu, and N. Zhao. Conservative front
tracking with improved accuracy. SIAMJNA, 41:19261947, 2003.
[25] J. Glimm, X.-L. Li, Y.-J. Liu, and N. Zhao. Conservative front tracking
and level set algorithms. Proc. National Academy of Sci., 98:1419814201,
2001.
[26] J. Glimm, D. Marchesin, and O. McBryan. Subgrid resolution of fluid
discontinuities II. J. Comput. Phys., 37:336354, 1980.
[27] J. Glimm, D. Marchesin, and O. McBryan. A numerical method for two
phase flow with an unstable interface. J. Comput. Phys., 39:179200,
1981.
[28] J. Glimm and O. McBryan. A computational model for interfaces. Adv.
Appl. Math., 6:422435, 1985.
[29] A. Harten and J. Hyman. Self-adjusting grid methods for one-dimensional
hyperbolic conservation laws. Rep. No. LA-9105, Los Alamos National
Laboratory, Los Alamos, NM, 1981.
[30] 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.
[31] R. L. Holmes, B. Fryxell, M. Gittings, J. W. Grove, G. Dimonte,
M. Schneider, D. H. Sharp, A. Velikovich, R. P. Weaver, and Q. Zhang.
Richtmyer-Meshkov instability growth: Experiment, simulation, and theory. J. Fluid Mech., 389:5579, 1999. LA-UR-97-2606.
65
[32] M. M. Koochesfahani and P. E. Dimotakis. Mixing and chemical reactions

in a turbulent liquid layer. J. Fluid Mech., 170:83112, 1986.
[33] H. Lee, H. Jin, Y. Yu, and J. Glimm. On validation of turbulent mixing
simulations of Rayleigh-Taylor mixing. Phys. Fluids, 20:18, 2008. Stony
Brook University Preprint SUNYSB-AMS-07-03.
[34] H. Lim, Y. Yu, J. Glimm, X.-L. Li, and D. H. Sharp. Chaos, transport,
and mesh convergence for fluid mixing. Acta Mathematicae Applicatae
Sinica, 24:355368, 2008. Stony Brook University Preprint SUNYSBAMS-07-09 Los Alamos National Laboratory preprint number LA-UR08-0068.
[35] H. Lim, Y. Yu, H. Jin, D. Kim, H. Lee, J. Glimm, X.-L. Li, and D. H.
Sharp. Multi scale models for fluid mixing. Compu. Methods Appl. Mech.
Engrg., 197:34353444, 2008. Stony Brook University Preprint SUNYSBAMS-07-05.
[36] J.-J. Liu, J. Glimm, and X.-L. Li. A conservative front tracking method.
In F. Asakura, H. Aiso, S. Kawashima, Matsumura A, S. Nishibata, and
K. Nishihara, editors, Hyperbolic Problems: Theory, Numerics, and Applications, pages 5762. Yokohama Publishers, Osaka, Japan, 2006.
[37] Jinjie Liu, Hyun-Kyun Lim, James Glimm, and Xiaolin Li. A conservative
front tracking method in N-dimensions. J. of Sci. Comp., 31:213236,
2007. Stony Brook University preprint number SUNYSB-AMS-06-04.
[38] X. F. Liu, Y. H. Li, J. Glimm, and X. L. Li. A front tracking algorithm
for limited mass diffusion. J. of Comp. Phys., 222:644653, 2007. Stony
Brook University preprint number SUNYSB-AMS-06-01.
[39] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3D
surface construction algorithm. Computer Graphics, 21(4):163169, 1987.
[40] T. Ma. Large eddy simulation of variable density flows. Ph.d. thesis,
University of Maryland, 2006.
[41] T. O. Masser. Breaking Temperature Equilibrium in Mixed Cell Hydrodynamics. Ph.d. thesis, State University of New York at Stony Brook,
2007.
[42] 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.
66
[43] A. S. Monin and A. M. Yaglom. Statistical Fluid Mechanics: Mechanics

of Turbulence. MIT Press, Cambridge, MA, 1971.
[44] R. Pember, John Bell, Phillip Colella, William Cruchfield, and Michael
Welcome. An adaptive cartesian grid method for unsteady compressible
flow in irregular regions. J. Computational Phys., 120:278304, 1995.
[45] R. Di Perna. Convergence of approximate solutions to conservation laws.
Arch. Rational Mech. Anal., 82:2770, 1983.
[46] R. Di Perna. Compensated compactness and general systems of conservation laws. Trans. Amer. Math. Soc., 292:383420, 1985.
[47] Heintz Pitsch. Large-eddy simulation of turbulent combustion. Annual
Rev. Fluid Mech., 38:453482, 2006.
[48] A. Polyanin. Hydrodynamics, Mass and Heat Transfer in Chemical Engineering, volume 14 of Topics in Chemical Engineering. CRC Press, 2002.
[49] G. C. Pomraning. Linear kinetic theory and particle transport in stochastic mixtures, volume 7 of Series on Advances in Mathematics for Applied
Sciences. World Scientific, Singapore, 1991.
[50] G. C. Pomraning. Transport theory in discrete stochastic mixtures. Advances in Nuclear Science and Technology, 24:4793, 1996.
[51] R. Richtmyer and K. Morton. Difference Methods for Initial Value Problems. Interscience, New York, second edition, 1967.
[52] H. F. Robey. The effect of viscosity and mass diffusion in hydrodynamically unstable plasma flows. Phys. of Plasmas, 11:41234133, 2004.
[53] D. H. Sharp. An overview of Rayleigh-Taylor instability. Physica D,
12:318, 1984.
[54] B. Swartz and B. Wendroff. Aztec: A front tracking code based on Godunovs method. Appl. Num. Math., 2:385397, 1986.
[55] C. Thomas, S. Kumar, G. Orlicz, and K. Prestridge. An experimental
investigation of mixing mechanisms in shock-accelerated flow. J. Fluid
Mech., 611:131150, 2008.
[56] Forman Williams. Combustion Theory. Addison-Wesley Co., Reading,
1965.
67
[57] P. Woodward and P. Colella. Numerical simulation of two-dimensional

fluid flows with strong shocks. J. Comput. Phys., 54:115, 1984.
[58] D. L. Youngs. Three-dimensional numerical simulation of turbulent mixing by Rayleigh-Taylor instability. Phys. Fluids A, 3:13121319, 1991.
[59] Y. Yu, M. Zhao, T. Lee, N. Pestieau, W. Bo, J. Glimm, and J. W. Grove.
Uncertainty quantification for chaotic computational fluid dynamics. J.
Comput. Phys., 217:200216, 2006. Stony Brook Preprint number SBAMS-05-16 and LANL preprint number LA-UR-05-6212.
68

Out 3

Uploaded by

Copyright:

Available Formats

Out 3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Out 3

Uploaded by

Copyright:

Available Formats

Numerical modeling in turbulent mixing flows

Applied Mathematics and Statistics

UMI Number: 3401712

All rights reserved

Stony Brook University

This dissertation is accepted by the Graduate School.

Abstract of the Dissertation

Overview and Motivation . . . . . . . . . . . . . . . . . . . . .

Front Tracking Method . . . . . . . . . . . . . . . . . . . . . .

Front Tracking Algorithm . . . . . . . . . . . . . . . .

Conservative Front Tracking Algorithm . . . . . . . . .

The Governing Equations . . . . . . . . . . . . . . . . . . . .

Numerical Development for LES . . . . . . . . . . . . . . . . .

2 Developments for LES with turbulence models

A Subgrid Scale Dynamic Model for the Momentum

A Subgrid Scale Dynamic Model for the Energy Equation 23

A Subgrid Scale Dynamic Model for the Concentration

Test Problem Setup . . . . . . . . . . . . . . . . . . . . . . . .

Correlation and Diffusion Length Scales . . . . . . . . . . . . .

Statistical Fluctuation vs. Mesh Convergence . . . . . . . . .

3 The Flow Instability Problem

4 Mesh Convergence for Macro Observables

5 Convergence of Micro Observables

6 Mathematical Existence and Uniqueness Theories

A sample of three types of interface representations. Upper-left:

Space time (r, t) contours of the primary waves, as detected by

Left: plot of C /x vs. Kmesh for a range of mesh levels and

The ratios (D + Dt )/C and (T + Tt )/C vs. K for several

Plot of (T ) vs. T . Cases l, g, p (rows 1 to 3), with Re increasing,

Plot of T pdf. Cases l, g, p (rows 1 to 3), with Re increasing, left

Radial dependence for hwi for the case l, Re 6000. . . . . . .

Transport coefficients considered in this paper, arranged in order

Ratios of dimensionless molecular to turbulent Schmidt and Prandtl

Mesh errors for (T ), defined by the weighted L1 norm (5.4).

I would like to express my profound gratitude to my advisor, Professor

Throughout my academic career, the constant support of my parents has

Overview and Motivation

Front Tracking Method

mary difficulty for numerical simulations. Commonly used numerical schemes

Front Tracking Algorithm

The front tracking method in the FronTier is performed by two main

In the interface propagation, the interface points are first propagated

Conservative Front Tracking Algorithm

Consider a system of hyperbolic conservation laws in N spatial dimension

By the divergence theorem, we have

(U, F (U)) ndS = 0 .

Figure 1.1: A sample of three types of interface representations. Upper-left:

(U, F (U)) ndS ,

U(x1 , ..., xN , tn )dx1 ...dxN

U(x1 , ..., xN , tn+1 )dx1 ...dxN

Eq. (1.4) degenerates into a static finite volume scheme

the space-time control volume reconstruction. While the clipping of the

and fully conservative front tracking method.

Developments for LES with turbulence models

The Governing Equations

S is the strain rate tensor

defined in terms of the velocity v and for any matrix A = Aij ,

A related convergence measure, but applicable to the concentration and

where the subgrid scale (SGS) variables are the ij , qi