1
The Mathematics and Numerical Principles for Turbulent Mixing
James Glimm1,3
With thanks to:
Wurigen Bo1, Baolian Cheng2, Jian Du7, Bryan Fix1,Erwin George4, John Grove2, Xicheng Jia1, Hyeonseong Jin5, T. Kaman1, Dongyung Kim1, Hyun-Kyung Lim1, Xaolin Li1, Yuanhua Li1, Xinfeng Liu6, Xingtao Liu1, Thomas Masser1,2, Roman Samulyak3, David H. Sharp2, Justin Iwerks1,Yan Yu1
1. SUNY at Stony Brook2. Los Alamos National Laboratory3. Brookhaven National Laboratory4. Warwick University, UK5. Cheju University, Korea6. University of Southern California7. University of South Carolina
2
Hyperbolic Conservation Laws
( ) 0tU F U Status of existence theory:
1D and small data: Perturbation expansion in wave interactions
existence, uniqueness,smooth dependence on data(G., Liu, Brezan, others)
1D and 2x2 system: Existence (Diperna, Ding, Chen, others)Measure valued solutions, then shown to beclassical weak solutions
2D: special solutions only
3
Nature of Solutions (1D) Solutions are typically discontinuous Shock waves form Solution space is functions of bounded
variation and L_infty Solutions are interpreted in the weak
sense, as distributions
4
Compensated Compactness Proofs (for 1D, large data, 2x2 systems)
First establish existence of solutions in a very weak space of measure valued distributions
Then (more difficult) show that such a weak solution is a classical weak solution, as a distribution:
Theorem (Gangbo and Westdickenberg): For 2D, 3D: only first step (measured valued solution) Isentropic equations Comm. PDEs (In press) Limit may not proved to satisfy the original equation
( ) 0 all test functions tU F U
5
Obstacles to extensions Some 3x3 and larger systems are unstable in
BV norm (but perhaps these are not physically motivated)
In 2D, even for gas dynamics, special solutions can be unbounded. Also BV is unstable.
Nonuniqueness, 2D: Scheffer (1989), Snirelman DeLellis and Szelkelyhidi (Archives) Volker Elling (numerical) Lopez, Lopez, Lowengrub and Zheng (numerical)
6
Numerical implications: Standard view Typically numerical solutions appear to
be convergent Problems with existence seems to be a
strictly mathematical concern This point of view is incorrect as
mathematics and as physics This lecture: to identify and cure
problems as mathematics and as physics
7
Implications, continued For many problems, physical regularization (viscosity, mass
diffusion, etc.) is very small. Even if included in the simulation, the effects are under resolved.
Thus effects of physical regularization are dominated by numerical effects such as numerical mass diffusion.
Large effort in V&V = Verification and Validation Verification is proof that the numerical solutions are bona fida
solutions of the mathematical equations. This step fails if mathematical solutions are nonunique or discontinuous
in dependence on initial conditions (loss of well posedness) or fail to exist or if the nonuniqueness etc. is resolved numerically by artifacts of algorithm
Validation is agreement with physical experiments In practice, validation fails dramatically for turbulent mixing
(Rayleigh-Taylor instability).
8
Chaotic Mixing: A challenge to the standard view Solutions are unstable on all length
scales Under mesh refinement, new structures
emerge In this sense there is no convergence Optimistically, we hope that the large scale
structures converge and the new small scale ones that emerge under mesh refinement will not influence the large scale ones
This hope is partially correct and partially not
9
Failure of the standard view Turbulent mixing and turbulent
combustion Atomic or molecular level mixing requires a
new length scale (the atoms or molecules) And a change in the laws of physics at
these length scales or above. New terms in the conservation laws, to
express mass diffusion, viscosity, heat conduction:( )tU F U D U
10
( )tU F U D U Even in 1D, and in the limitdistinct solutions occur (Smoller, Conley)
Ratios of diffusion terms at the root of issue
For combustion, dimensionless ratios in D influence mixing properties and combustion rates, thus the global dynamics (Schmidt and Prandtl numbers). In this sense the small scales can easily affect the large ones.
For turbulent mixing, parabolic transport terms can affect solution, especially the ratios of transport terms Density differences drive instabilities, and so mass diffusion
diminishes the instability. Viscosity limits the complexity of the flow and hence the amount of the interfacial mass diffusion. In the limit D 0, the Schmidt number (viscosity/mass diffusion) determines the net amount of mass diffusion and hence the RT instability growth rate.
0D
Parabolic Conservation Laws
11
Requirements for simulation of turbulent combustion Averaged flow velocities and pressures as a
function of x,y,z,t Joint probability distributions of the
concentrations of species (fuel, oxidizer) and temperature, as a function of x,y,z,t
This depends on the diffusion matrix D and if the calculation is under resolved, it depends on the numerical code and the numerical D, not the physical one
12
Mathematical implications for hyperbolic conservation laws The ultra weak solutions as measure valued
distributions (PDFs) depending on x,y,z,t is exactly the framework of compensated compactness
Existence proofs in this framework should be easier than for classical (weak) solutions
Uniqueness, which typically fails for such weak solution methods, should fail, and is not correct (in the hyperbolic setting), neither as mathematics nor as physics.
On numerical grounds, appears to correctly express the physics of unregularized solutions.
13
Numerical example: Richtmyer-Meshkov (shock accelerated) turbulent mixing
Circular shock starts at outer edge of (1/2) circle.
Moves inward, crosses a perturbed circular density difference
Reaches origin, reflects, recrosses the density difference
Highly unstable flow Illustrates all previous points
14
Numerical example: Primary breakup of jet (diesel fuel); Jet in cross flow (jet engine) Liquid leaves narrow nozzle at speed near Mach 1.
Breaks up rapidly into small droplets. Breakup is essential feature of flow to predict
combustion rates. Burning is mainly film burning, and so is rate limited by surface area of droplets.
For scram jet, fuel leaves combustion chamber in 1 millisec, and must complete combustion before that.
Flows are only stabilized by transport (viscosity, mass diffusion or surface tension and heat conduction.
Illustrates main points for chaotic mixing.
15
Numerical example:Chemical fuel separation
in a mixer/contactor High shear rate in rotating device (Couette flow)
produces high levels of Kelvin-Helmholtz instability Interface between immiscible fluids (oil and water
based) produces fine scale droplets. Chemical reactions occur on the interface and are
thus interface area limited. Prediction of total droplet area and distribution of
droplet sizes and shapes is needed to predict the chemical reaction rates.
Flow is stabilized by surface tension (Weber number).
Illustrates main points for chaotic flow.
16
Numerical Example: Rayleigh-Taylor Turbulent Mixing
Steady acceleration of a density difference
Unstable if acceleration points from heavy to light Water over air
Penetration distance h of light fluid into heavy (bubbles) is a simple measure of mixing rate. 2/h Agt
17
2D Chaotic Solutions: Shock implosion of perturbed interface—4 grid levels
18
CircularRM instabilityInitial (left)and afterreshock (right) densityplots. Upper andlower insertsshow enlargeddetails of flow.
19
Definitions for Turbulent Mixing An observable of the flow is SENSITIVE if it
shows dependence on numerical algorithms and/or on physical modeling (transport).
MACRO observables: edges of mixing zones, trajectories of leading shock waves, mean densities, velocities of each fluid species
MICRO observables: probability density functions for temperature and for species concentration; properties dependent on atomic mixing
20
Explanation Macro observables describe the
average mixing properties of the large scale flow
Micro observables describe the atomic level mixing properties
21
Sensitive and Insensitive Observables
SensitiveInsensitiveRichtmyer-Meshkov
SensitiveSensitiveRayleigh-Taylor
MicroMacro
RM has a highly unstable interface, in theabsence of regularization, diverging as 1/delta x.
This divergent interface length or area causes sensitivity to atomic scale observables.
22
Simplistic Error Analysis for Micro Observables in Chaotic Flow
1
12
11 2 1 2
Interface length (area)
Interface length (area) =
1
ERROR C x
C x
ERROR C x C x C C O
Error as results from numerical or physical modeling, e.g.numerical mass diffusion or ideal vs. physical transport coefficients
23
Chaotic dependence of interface length on mesh: Unregularized simulation
Surface Length/Area
vs. time
Surface Length
Areax
24
Verification (Macro Observables) A mesh convergence study Analyze heavy fluid density Errors are of two types
Position errors due to errors in locations of shock waves Density errors in the smooth regions between the shock
waves Large reduction in error variance results from
separating these types of errors Some statistical averages performed to achieve
convergence Convergence achieved in the sense of mean
quantities
25
Heavy fluid relative density errors, at t = 60
Convergence of mean density of heavy fluid
26
Convergence of wave trajectories
Kolmogorov length =
= size of smallest eddy
Kolmogorov length / x
1: DNSKmesh
Kmesh
<- change in physics ->
27
Validation: Agreement betweensimulation and experiment Rayleigh-Taylor instability
Classical hydro instability, acceleration driven Three simulation studies based on Front
Tracking show agreement with experiment Immiscible fluids (surface tension is regularizer) Miscible fluids high Schmidt number (viscosity and
initially mass diffused layer are regularizers) Miscible fluids (helium-air) (mass diffusion is
regularizer) All three cases give essentially perfect agreement
to experiment on overall mixing rate.
28
Simulation compared to experiment:Rayleigh-Taylor mixing with small levels of mass diffusion (large Schmidt number).
Plot of h = penetration distance of light fluid into heavy.
Physical values of viscosity and initialmass diffusion are used.
The two simulations arequarter and half of physical problem.
29
Comparison of simulations and experiment (Validation): For Rayleigh-Taylor unstable mixingImmiscible fluids (surface tension)
growth rate
for mixing zone
dimensionless
surface tension
30
Other RT Simulation Studies 50 year old problem Most simulations disagree with experiment
by factor of 2 on overall mixing rate
A few agree with experiment Front Tracking (control numerical mass
diffusion) Particle methods (control numerical mass
diffusion) Mueschke-Andrews-Schilling (control numerical
mass diffusion and use experimental initial conditions)
2/h Agt
31
Comparison of FronTier and RAGEfor 2D RM instability
T. Masser: Macro variables not sensitive Micro: Temperature is sensitive
Stony Brook: Macro variables: not sensitive Interface length divergent Micro: Mixture concentrations sensitive
32
Major Temperature Differences at Reshock
At reshock the fingers of tin are heated to a much higher temperature in the FronTier simulation than the corresponding fingers in the RAGE simulation.
There are at least three possible mechanisms that might be responsible.
Velocity shear in FronTier missing in RAGE
Thermal and Mass diffusion at the interface in RAGE
Differences in the hyperbolic solver After reshock FronTier continues to have
a significantly higher maximum temperature.
33
DNS (regularized simulation) convergence of interfacelambda_C = avg. distance to exit a phaselambda_Cmesh = lambda_C/mesh sizelambda_K = Kolmogorov length (smallest eddies)
LHS of left frame shows uniform behavior in mesh units, independent of mesh and physics, for LES regime.RHS of right frame shows uniform behavior relative to meshfor fixed physics, ie mesh convergence for DNS regime.
34
Atomic Scale Mixing Properties Atomic scale mixing properties are sensitive
to physical modeling and to numerical methods unless fully resolved (direct numerical simulation = DNS)
Correct simulation for both micro and macro variables: use sub grid models (large eddy simulation = LES)
LES modify equations to compensate for physics occurring on small scales (below the grid size) but not present in the computation.
35
Combine strengths from two numerical traditions Capturing likes steep gradients, rapid time
scales Tracking is an extreme version of this idea Often, no subgrid model and so not physically
accurate for under resolved (LES) simulations Turbulence models often use smooth
solutions, slow time scales with significant levels of physical mass diffusion Often, too many zones to transit through a
concentration gradient Best part of two ideas combined in present
study
36
Subgrid models for turbulence, etc. Typical equations have the
form Averaged equations:
( )tU F U U
( )
( ) ( )
( ) ( ) ( )
t
SGS
U F U U
F U F U
F U F U F U
is the subgrid scale model and corrects for grid errorsSGSF
37
Subgrid models for turbulent flow
turbulent
turbulent
(key modeling step)
SGS
t SGS
SGS
t
F U F U F U
U F U U F U
F U U
U F U U
;
38
Subgrid Scale Models (Moin et al.) No free (adjustable) parameters in the SGS terms Parameters are found dynamically from the
simulation itself After computing at level Delta x, average solution
onto coarser mesh. On coarse mesh, the SGS terms are computed two ways: Directly as on the fine mesh with a formula Indirectly, by averaging the closure terms onto the coarse
grid. Identity of two determinations for SGS terms becomes an
equation for the coefficient, otherwise missing. Assume: coefficient has a known relation to Delta x and
otherwise is determined by an asymptotic coefficient. Thus on a fine LES grid the coefficient is known by above algorithm.
39
Typical atomic level observable Mean chemical reaction rate (no
subgrid model needed for computation of w)
/1 2
1 2
1 2
/1 2
Activation Temperature
mass fraction of species
; defined at fixed
( ) probability distribution for
( ) mean reaction rate
AC
AC
AC
T T
AC
i
T
T TT
T
w f f e
T
f i
f fT
f f
d T T
w f f e d T
L
40
Re = 300. Theta(T) vs. T (left); Pdf for T (right) 1 2
1 2
f f
f f
41
Re = 3000. Theta(T) vs. T (left); Pdf for T (right)
42
Re = 300k. Theta(T) vs. T (left); Pdf for T (right)
43
Kolmogorov-Smirnov Metric for comparison of PDFs Sup norm of integral of PDF
differences PDF data is noisy and the smoothing
in the K-S metric is needed to obtain coherent results
1 2 1 2|| || || ( ) ( ) ||x
K Sp p p y p y dy
44
Convergence properties for reaction rate pdf for
Conclusion: numerical convergence of chemical reaction rates, using LES SGS models for high Schmidt number flows
1 2 const. exp( / )ACw f f T T
Errors for pdf for reaction rate w, compare coarse to fine, medium to fine and relative fluctuations in coarse grid
0.250.030.09300K
0.490.040.493K
0.240.030.04300
fluct. cm to f
c to fRe
45
Convergence of w pdf Not just mean converges Moments to all order, and full
distribution converges
46
Microphysics to study combustion:Primary breakup of fuel jet injection to engineParameters from diesel engine
Above: no cavitation bubbles. Below: with cavitation
47
Comparison to experiment:simulation with and without cavitation bubbles
Velocity of jet tip Mass flux through observation window asa measure of jet spreading
Reynolds numbervs. time
Base case: specification of numerical parameters for insertion of cavitation bubbles
48
3D Simulation
49
Comparison of simulation and experiment (Argon National Laboratory)
50
Simplified engine geometry
51
Jet in Cross Flow
Mach 1 cross flow with a2D (planar) jet of dieselfuel.
52
Conclusions Chaotic flow simulations are sensitive to numerical and physical
modeling Several examples of natural physical problems RM mixing: Temperature, concentration and chemical reaction
rate PDFs are sensitive to transport, to numerical algorithms (under resolved) and are convergent with use of SGS model (no adjustable parameters).
RM mixing: mesh convergence including microphysical variables (verification)
RT mixing: agreement of simulation with experiment (validation)
Solutions as measure valued distributions from the compensated compactness theory provides a useful framework for interpretation of simulations.
Simulations suggest that this framework might be a basis for 2D/3D existence proofs.
53
Thank YouSmiling Face: FronTier art simulation
Courtesy of Y. H. Zhao