materials process design and control laboratory a stabilized stochastic finite element second-order...
TRANSCRIPT
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
A stabilized stochastic finite element second-order
projection method for modeling natural convection in random
porous media
Materials Process Design and Control LaboratorySibley School of Mechanical and Aerospace Engineering
188 Frank H. T. Rhodes HallCornell University
Ithaca, NY 14853-3801
Email: [email protected]: http://mpdc.mae.cornell.edu
Xiang Ma and Nicholas Zabaras
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Outline of the presentation
1. Motivation and problem definition
2. A stabilized second-order projection method
3. Representation of stochastic processes
4. Stochastic finite element method formulation
5. Numerical examples
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Transport in heterogeneous media
- Thermal and fluid transport in heterogeneous media are ubiquitous, e.g. solidification
- Range from large scale systems (geothermal systems) to the small scale
- Complex phenomena
- How to represent complex structures?
- How to make them tractable?
- Are simulations believable?
- How does uncertainty propagate through them?
To apply physical processes on these heterogeneous systems
- worst case scenarios
- variations on physical properties
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Problem of interest
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
= 1
u = v =0
= 0
u = v =0
= 0, u = v =0n
= 0, u = v =0n
Natural convection in random porous media
Deterministic governing equation:
Pr -Prandtl number
Ra- Raleigh number
Da- Darcy number
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Problem of interest
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
= 1
u = v =0
= 0
u = v =0
= 0, u = v =0n
= 0, u = v =0n
( )w
Natural convection in random porous media
Modeling porosity as a Gaussian random field
( )w
Stochastic governing equation:
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Problem of interest: Some difficulties
solid Mushy zone
q
liquid ~ 10-2 m
This model is important in alloy solidification process. To understand the uncertainty propagation in this process can help understand error in the experiment results.
A SUPG/PSPG formulation has been proposed to solve this complex problem. ( Deep & Zabaras, 2004). However, due to its coupling between velocity and pressure, it resulted in a very large linear system.
very hard to extend to stochastic formulation due to the high degree of freedom. Especially, when dealing with real porous media, it results in a very high stochastic dimension.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Problem of interest: Some difficulties and solutions
The projection method for the incompressible Navier-Stokes equations, also known as the fractional step method or operator splitting method, has attracted widespread popularity.
The reason for this lies on the decoupling of the velocity and pressure computation.
It was first introduced by Chorin, as the first-order projection method ( also called non-incremental pressure-correction method). Later, it was extended to the second-order scheme ( also called incremental pressure-correction scheme ) in which part of the pressure gradient is kept in the momentum equation.
Le Maitre et.al(2002) have developed a stochastic projection method to model natural convection in a closed cavity based on first order method.
It is really helpful when solving the high stochastic dimension problem.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Let us review first order method is
Problem of interest: Some difficulties and solutions
1/ 2 1/ 21
( ) ( ) 0n n nu u N u udt
1 1/ 2 1
1
1( ) 0
0
n n n
n
u u pdt
u
2 1 1/ 21n np udt
1 1/ 2 1
1
1 1/ 2
1( ) 0
0
0
n n n
T n
n T n
M U U GPdt
G U
dtLP G U
1 1 1( ) 0T n T nG U dt L G M G P
solve for intermediate velocity:
projection step: eliminate : 1nu
Once a finite element discretization is performed, the matrix form is
eliminate : 1nu
The term may be understood as the difference between two discrete Laplacian operators computed in difference manner. This matrix turns out to be positive semi-definite, which increases the stability of the numerical method, thus allows equal-order interpolation.
1TL G M G
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Problem of interest: Some difficulties and solutions
For the porous media problem considered here, due to the porosity dependence of the pressure gradient term in the momentum equation, we cannot omit the pressure term as is the case in the first-order projection method. We need to use second-order scheme.
porosity dependence
However, if using FEM, second-order projection method is known that it exhibits spatial oscillation in the pressure field if not using a mixed finite element formulation for velocity and pressure. (J.L.Guermond et al. 1998)
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Problem of interest: Some difficulties and solutions
In order to utilize the advantage of the incremental projection method which retains the optimal space approximation property of the finite element and allows equal-order finite element interpolation, Codina and co-workers developed a pressure stabilized finite element second-order
projection formulation for the incompressible Navier-Stokes equation. (Codina et al.2000)
The method consists in adding to the incompressibility equation the divergence of the difference between the pressure gradient and its projection onto the velocity space, both multiplied by algorithmic parameters defined element-wise.
pressure field in a lid-driven cavity problem with stabilization (left) and without stabilization (right).
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Pressure stabilized formulation
We take these parameters as
where is the viscosity, is the local size of the element , is the local velocity and and are algorithmic constants, that we take as and for linear elements.
Kh K hu
1c 2c 1 4c 2 2c
Accordingly, the continuity equation is modified as follows:
where is the stabilized parameter as discussed before and the new auxiliary variable is the projection of the pressure gradient onto the velocity space.
p
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Pressure stabilized formulation
A matrix version of this form is
( , ) ( , ) ( , ) 0,
( , ) ( , ) 0 h h h h h h h
h h h h h h
p q q q v q Q
p v v v V
0 0
0
T TLP G G U
GP M
weak form:
Eliminating from equation, it is found that
Let’s see how it works:
10( ) 0T TL G M G P G U
which is similar to 1 1 1( ) 0T n T nG U dt L G M G P
So this stabilized method mimics the stable effect of the first-order method. Thus it allows the equal-order interpolation and stabilize the pressure.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stabilized second-order projection method
Step 1: Solve for the intermediate velocity in the momentum equation: 1/ 2nv
(1)
Fully decoupled, can be solved one by one.
Step 2: Projection step
Step 2: update step
(2)
(3)
(4)
(5)
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
A numerical example
Free fluid
Porous Medium
= 1
u = v =0
= 0
u = v =0
= 0, u = v =0n
= 0, u = v =0n
l 1 w
The computation region is a square domain.
The dimensionless length is taken as
The wall porosity is taken as 0.4. The porosity increases linearly from 0.4 at the wall to 1.0 (pure liquid) at the core.
The Rayleigh number is , Prandtl number is 1.0 and the Darcy number is
The problem was solved with a mesh discretization
and the time step was chosen as
The simulation was run up to non-dimensional time 1.0.
[0,1] [0,1]
l 0.3l
w
61 1076.665 10
50 50 55 10
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
A numerical example
-9
-8
-7
-6-5
-4
-4
-3
-3
-1
-1
1
2
2
4
-5.48
-1.73
-1.73
-0.0
9
-0.09
-0.0
9
-0.09
-0.0
1
-0.01
-0.01
Streamline Isothermal
First-order solution
Second-order solution
The Streamline pattern is symmetric
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
A numerical example
65000055000045000035000025000015000050000
806040200
-20-40-60
180160140120100806040200
-20
180000160000140000120000100000800006000040000200000
403020100
-10-20-30-40
50403020100
-10-20-30-40-50
First-order solution
Second-order
solution
u velocity v velocity Pressure
The velocities are mainly distributed in the free fluid region
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Representation of stochastic processes
For special kinds of stochastic processes that have finite variance-covariance function, we have mean-square convergent expansions
Series expansions
Known covariance function
Unknown covariance function
Best approximation in mean-square sense
Useful typically for input uncertainty modeling
Can yield exponentially convergent expansions
Used typically for output uncertainty modeling
Karhunen-Loeve Generalized polynomial chaos
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Karhunen-Loeve expansion
1
( , , ) ( , ) ( ) ( )i i ii
X x t X x t f x
Stochastic process
Mean function
ON random variablesDeterministic functions
is the mean of the process and let denote its covariance function, which is needed to be known a priori.
is a set of i.i.d. random variables, whose distribution depends on the type of stochastic process.
and are the eigenvalues and eigenvectors of the covariance kernel. That is ,they are the solution of the integral equation:
In practice, we truncate KLE to first M terms.
( , )X tx 1 2( , )R x x
1{ }i i
i ( )if x
1 2 1 1 2( , ) ( ) ( )i i iDR x x f x dx f x
1( , , ) fn( , , , , )NX x t x t
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Generalized polynomial chaos
Generalized polynomial chaos expansion is used to represent the stochastic output in terms of the input
0
( , , ) ( , ) (ξ( ))i ii
Z x t Z x t
Stochastic output
Askey polynomials in inputDeterministic functions
Stochastic input
1( , , ) fn( , , , , )NX x t x t
Askey polynomials ~ type of input stochastic process
Usually, Hermite (Gaussian), Legendre (Uniform), Jacobi etc.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Generalized polynomial chaos
The polynomial chaos forms a complete orthogonal basis in the of the Gaussian random variables, i.e.
2L
2
ii j ij
where denotes the expectation or ensemble average. It is the inner product in the Hilbert space of the Gaussian random variables
,
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )f g f g dP f g W d ξ ξ ξ ξ ξ ξ ξ ξ ξ
where the weighting function has the form of the multi-dimensional independent Gaussian probability distribution with unit variance.
( )W ξ
1 dimension:21
21
1( )
2W e
2 dimensions:2 21 2
1 1
2 22 1 2 1 1 1 2
1( ) ( ) ( )
2W W W e e
n dimensions:2 21
1( )
21 1 1 1
1( ) ( ) ( )
(2 )
n
n n n nW W W e
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
PC computation
PC computation (B. Debusschere et al. 2004, Ma & Zabaras, 2007):
Quadratic product:
Consider two random variables, and , with their respective GPCE:
We want to find the coefficients of :
The coefficient are obtained with a Galerkin projection, which minimizes the error of the resulting PC representation within the space spanned by the basis function up to order P.
kc
kc
a b
c
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
PC computation
Quadratic product: Thus we can obtain:
with
A similar procedure could also be used to determine the product of
three stochastic variables d abc
with
Triplet product:
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
PC computation
Sampling approach for general nonpolynomial function evaluations:
We consider a general nonlinear function , where is
We need to express this function as
Thus we obtain
( , )f x
Using the same method as before (Galerkin projection), we have
High dimension integration, use
Latin Hypercube sampling.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic Finite Element Formulation
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Since there are nonlinear functions of in the governing equation, we need to first express them in the polynomial basis using the method discussed before:
( , ) x
0
1ˆ ( )
P
i ii
ξ2
20
(1 )( )
P
i ii
ξ 2
0
1( )
P
i ii
ξ
Since the input uncertainty is taken as a Gaussian random field, we use Hermite polynomials to represent the solution:
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic Finite Element Formulation
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Substitute the above equation into Eqs (1)-(5)and then performing a Galerkin projection of each equation by , k
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic Finite Element Formulation
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
In the non-linear drag term
we assume that the magnitude of the velocity is determined by the
mean velocity From the pressure update equation
the stabilized parameter is determined by the kth coefficient of the spectral expansion. So we denote it as to emphasize that it is a function of the k-th coefficient
nv
0nv
k
kv
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic Finite Element Formulation
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
It is time consuming to evaluate the fourth order product term
directly using the method discussed before. To simplify this calculation, we introduce an auxiliary random variable as follows:
such that
So our term of interest can be reduced to
This form is now easier to calculate, since it only evolves third-order product terms, which is pre-computed.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Some computational details
1. All calculations are performed using numerical algorithms provided in PETSc. (Portable, Extensible Toolkit for Scientific computation). The default Krylov solver was used for the linear solver).
2. For the numerical computation of KLE, the SLEPc is used. ( Scalable Library for Eigenvalue Problem Computation). It is based on the PETSc data structure and it is highly parallel.
3. In order to further reduce the computational cost, we solve each expansion coefficient individually. That means we split the systems of equations into (P+1)(3d+2) deterministic scalar equations.
4. The computation is performed using 30 nodes on a local Linux cluster.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media: large dimension
= 1
u = v =0
= 0, u = v =0n
= 0, u = v =0n
( )w = 0
u = v =0
The porosity is modeled as a Gaussian random filed. And the correlation function is extracted from a real porous media – sandstone 1. The mean is 0.6.
The computation region is a square domain.
The Rayleigh number is , Prandtl number is 1.0 and the Darcy number is
The problem was solved with a mesh discretization
and the time step was chosen as
The simulation was run up to non-dimensional time 1.0.
61 10
[0,1] [0,1]
76.665 10
30 30 31 10
Alloy solidification, thermal insulation, petroleum prospecting
Look at natural convection through a realistic sample of heterogeneous material
1. Reconstruction of random media using Monte Carlo methods, Manwat and Hilfer, Physical Review E. 59 (1999)
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0 10 20 30 40 50 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Material: Sandstone
Numerically computed
Eigen spectrum
Experimental correlation for the porosity of the sandstone.
Eigen spectrum is peaked. Requires large dimensions to accurately represent the stochastic space
KLE is truncated after first 6 terms.
So the stochastic dimension is 6.
Correlation function
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.0140.0130.0120.0110.010.0090.0080.0070.0060.005
0.010.0080.0060.0040.0020
-0.002-0.004-0.006-0.008-0.01
0.0080.0060.0040.0020
-0.002-0.004-0.006-0.008
Eigenmodes for a 2-D domain
Mode 1 Mode 2
Mode 4Mode 6
0.0060.0040.0020
-0.002-0.004-0.006
( )f x
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.6120.6080.6040.60.5960.5920.5880.5840.58
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.58
0.59
0.6
0.61
0.6120.6080.6040.60.5960.5920.5880.5840.58
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.58
0.59
0.6
0.61
Two realizations of the porosity fields with 6 terms:
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
We have already obtained the KLE of the Gaussian random fields.
1
( , ) ( , ) ( ) ( )M
i i ii
f
x x x
Now We want to express the following non-polynomial functions in terms of polynomial chaos.
0
1ˆ ( )
P
i ii
ξ2
20
(1 )( )
P
i ii
ξ 2
0
1( )
P
i ii
ξ
We can evaluate the coefficients of the expansions for the three non-linear functions using LHS. Then we can sample the calculated GPCE expansion to obtain the PDF.
The optimal order is determined such that the GPCE expansion can accurately represent the PDF of these non-linear functions.
We compare the results with the “Direct Sampling” approach, where the PDF of the functions is obtained by sampling form the standard normal distribution, then calculate the realization of each sample and plot the PDF.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
1.5 1.6 1.7 1.80
2
4
6
8
10
order 1order 2order 3Direct Sampling
P
roba
bili
ty
1
( )
0.2 0.3 0.4 0.5 0.6 0.70
1
2
3
4
5
6
7 order 1order 2order 3Direct Sampling
P
roba
bili
ty
2
2
(1 ( ))
( )
0.8 1 1.2 1.40
1
2
3
4 order 1order 2order 3Direct Sampling
P
roba
bili
ty
2
1 ( )
( )
A second-order GPCE is enough to capture
all of the input uncertainties
Thus, we choose a 6-dimension second order GPCE expansion to represent the solution process, which give a total of 28 modes.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
Mean
7531
-1-3-5-7
86420
-2-4-6-8
0.950.850.750.650.550.450.350.250.150.05
650055004500350025001500500
u v
p
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
1u2u 3u
4u 5u 6u
0.350.250.150.05
-0.05-0.15-0.25-0.35
0.180.140.10.060.02
-0.02-0.06-0.1-0.14
0.160.120.080.040
-0.04-0.08-0.12-0.16
0.090.070.050.030.01
-0.01-0.03-0.05-0.07-0.09
0.070.050.030.01
-0.01-0.03-0.05-0.07
0.070.050.030.01
-0.01-0.03-0.05-0.07
First order u velocity
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.40.30.20.10
-0.1-0.2-0.3-0.4
1v
0.10.060.02
-0.02-0.06-0.1-0.14
0.140.10.060.02
-0.02-0.06-0.1-0.14-0.18
2v 3vFirst order v velocity
0.10.080.060.040.020
-0.02-0.04-0.06-0.08-0.1
0.080.060.040.020
-0.02-0.04-0.06-0.08
0.080.060.040.020
-0.02-0.04-0.06-0.08
4v 5v 6v
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.0080.0060.0040.0020
-0.002-0.004-0.006-0.008
0.00060.0002
-0.0002-0.0006-0.001-0.0014-0.0018
0.0010
-0.001-0.002-0.003-0.004-0.005
0.00140.0010.00060.0002
-0.0002-0.0006-0.001-0.0014
0.00160.00120.00080.00040
-0.0004-0.0008-0.0012-0.0016
0.00060.00040.00020
-0.0002-0.0004-0.0006
1 2 3
4 5 6
First order Temperature
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.0050.0040.0030.0020.0010
-0.001-0.002-0.003
0.0010.00080.00060.00040.00020
-0.0002-0.0004-0.0006-0.0008-0.001-0.0012
0.00060.00040.00020
-0.0002-0.0004-0.0006
0.0050.0040.0030.0020.0010
-0.001-0.002-0.003-0.004
0.00060.00040.00020
-0.0002-0.0004-0.0006-0.0008
0.0010.00080.00060.00040.00020
-0.0002-0.0004-0.0006
6E-052E-05
-2E-05-6E-05-0.0001-0.00014-0.00018
3.5E-052.5E-051.5E-055E-06
-5E-06-1.5E-05-2.5E-05-3.5E-05
8E-066E-064E-062E-060
-2E-06-4E-06-6E-06-8E-06
9u
9v
9
14u
14
14v
25u
25
25v
Second order Modes
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
5 6 7 80
0.2
0.4
0.6
0.8
1
1.2
SSFEMMC
-0.5 0 0.50
0.5
1
1.5
2
2.5
3
SSFEMMC
velocityv
P
roba
bili
ty
velocityu
P
roba
bili
ty
P
roba
bili
ty
Temperature0.64 0.65 0.66 0.67
0
20
40
60
80
SSFEMMC
PDF at one point
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.380.340.30.260.220.180.140.10.060.02
0.460.420.380.340.30.260.220.180.140.10.060.02
0.00850.00750.00650.00550.00450.00350.00250.00150.0005
4642383430262218141062
Standard deviation
u v
p
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
7531
-1-3-5-7
0.950.850.750.650.550.450.350.250.150.05
0.950.850.750.650.550.450.350.250.150.05
7531
-1-3-5-7
0.950.850.750.650.550.450.350.250.150.05
7531
-1-3-5-7
GPCE
Collocation
Monte Carlo
Compare Mean
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Natural convection in heterogeneous media
0.380.340.30.260.220.180.140.10.060.02
0.00850.00750.00650.00550.00450.00350.00250.00150.0005
0.380.340.30.260.220.180.140.10.060.02
0.380.340.30.260.220.180.140.10.060.02
0.00850.00750.00650.00550.00450.00350.00250.00150.0005
0.00850.00750.00650.00550.00450.00350.00250.00150.0005
GPCE
Collocation
Monte Carlo
Compare standard deviation
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
3-D Natural convection in heterogeneous media
0.5950.59450.5940.59350.5930.59250.5920.59150.5910.59050.59
One realization of the random porosity random field in 3-D
contour
porosity iso-surface
porosity slice of the xz
plane at y=0.5
The definition is the same as 2-D, except here we consider a covariance function 2( ) 0.05 exp( )
10
rR r
KLE is truncated after two terms, so the stochastic dimension is 2.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
3-D Natural convection in heterogeneous media
1.4 1.6 1.8 2 2.2 2.40
0.5
1
1.5
2
2.5
3order 1order 2order 3order 4Direct Sampling
P
roba
bili
ty
1
( )
P
roba
bili
ty
2
2
(1 ( ))
( )
0.4 0.8 1.20
0.5
1
1.5
2
2.5
order 1order 2order 3order 4Direct Sampling
1 2 30
0.2
0.4
0.6
0.8
1
1.2
1.4
order 1order 2order 3order 4Direct Sampling
P
roba
bili
ty
2
1 ( )
( )
A second-order GPCE is not enough to capture all of the input uncertainties. At least, we need a third-order GPCE.
Thus, we choose a 2-dimension third order GPCE expansion to represent the solution process, which give a total of 10 modes.
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Mean slice of the xz plane at y=0.5
XY
Z
0.950.850.750.650.550.450.350.250.150.05
XY
Z
0.950.850.750.650.550.450.350.250.150.05
Collocation
GPCE
Left: GPCE
Right: Collocati
on
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Standard deviation slice of the xz plane at y=0.5
0.43 0.44 0.45 0.46 0.47 0.48 0.490
10
20
30
40
P
roba
bili
ty
Temperature-0.3 -0.2 -0.1 0
0
2
4
6
8
10
P
roba
bili
ty
velocity
u
GPCE
Collocation
Left: GPCE
Right: Colloca
tion
PDF at one
point
Materials Process Design and Control LaboratoryMaterials Process Design and Control Laboratory
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Conclusions
A second-order stabilized stochastic projection method was used to model uncertainty propagation in natural convection in random porous media.
For the problems examined, the computation cost of the SSFEM and sparse grid simulation were similar. The MC simulations were significantly more expensive. In the SSFEM, it was shown that it is rather easy to identify dominant stochastic models in the solution and investigate how the uncertainty propagates from porosity to the velocity and temperature random fields.
The key ingredient in the implementation of the algorithms presented here includes the development of a stochastic modeling library based on the GPCE formulation. This library includes computation tools for the implementation of the Askey polynomials, a parallel K-L expansion eigen solver, and a post-processing class for calculation of higher-order solution statistics such as standard deviation and probability density function.