Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
THE MULTI-LEVEL MONTE CARLO FINITE ELEMENT METHOD FOR A
STOCHASTIC
BRINKMAN PROBLEM
CLAUDE GITTELSON*, JUHO KÖNNÖ**, CHRISTOPH SCHWAB* AND ROLF STENBERG***
15 July 2011
1 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
* ETH Zürich, Seminar für angewandte Mathematik
** Wärtsilä Finland Oy, R&D/C&S
*** Aalto University, Department of Mathematics and Systems Analysis
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Outline
• The Brinkman model
• Stochastic permeability
• Karhunen-Loève expansion
• Main ideas of the MLMC method
• Requirements from the finite element method – Implementation
• Numerical results
15 July 2011
2 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
The Brinkman model
• Describes the flow of a viscous fluid in a porous medium – Cracks and flow channels in porous media
– Heat pipes, oil filters, composite resin infusion – Combination of the Darcy and Stokes models
15 July 2011
3 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Norms and variational spaces
• The nature of the problem changes when going to the Darcy limit
• The following norms are employed
15 July 2011
4 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Stokes Darcy
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
The weak form
• Find a velocity-pressure pair
15 July 2011
5 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
such that
in which
and
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Stochastic permeability
• The permeability tensor is assumed to be a random field
• A log-normal model is employed for the permeability
• Often only statistical data is available for the permeability
15 July 2011
6 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
• G is a Gaussian field and M 0 is positive definite
• M is integrable with respect to the probability measure
• Implies integrability of the velocity-pressure pair
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
K-L expansion
• The Gaussian field can be expanded as the Karhunen-Loève series
15 July 2011
7 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
• Next, the series is approximated by truncating after N terms Normalized eigenpairs of
the covariance operator
Normal random variables
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
K-L expansion
15 July 2011
8 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
• The following truncation estimate holds for the Gaussian field G
• The parameter s depends on the convergence speed of the eigenvalues of the covariance operator
• Holds for any
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Truncated log-normal field
• For the solution of the problem with the truncated permeability field it holds
• How to decide what N should be?
15 July 2011 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö 9
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
The MLMC ideology
• Multi level idea: balance the stochastic truncation error and discretization error
• Different eigenfunctions correspond to different „frequencies‟
– Sparse mesh is sufficient for low frequencies
• Traditional Monte Carlo with M samples
15 July 2011 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö 10
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
The MLMC method
• Based on the following telescoping sum property:
• Use different level of approximation on each level l
• “Compute successive increments, not the whole field”
15 July 2011 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö 11
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Error estimates
• The following bounds hold for a L-level method
• We have to balance the MC error compared to
– FEM errors
– Truncation error
15 July 2011 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö 12
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Number of samples and workload
• Assuming uniform mesh refinement we choose the truncation level and number of MC samples as
• Yields optimal order of convergence with respect to the polynomial degree
• Workload is
15 July 2011 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö 13
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Requirements towards FEM
• Telescoping sum property must hold for the discrete solution – In practice, nested FE spaces are required
• Problems with bubble degrees of freedom, e.g. MINI element
– Interpolation could be done, but too expensive
• Stabilization parameters might depend on stochastic data – Must be easily computable or estimated
• Evaluating the K-L series is a major computational effort – Low-order methods, reduced integration
15 July 2011
14 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Our approach to MLMC-FEM
• Stabilized equal order Stokes-based elements
• Mesh-dependent norm
• Residual stabilization
– Parameter easily computed for linear elements
15 July 2011 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö 15
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
FEM error estimates
• Estimates follow those presented by Juntunen & Stenberg (Calcolo 2009) for the Brinkman problem
– Nature of the problem changes numerically at t=h – Modified to include stochastic permeability
– Constants must be tracked carefully to assure integrability
15 July 2011
16 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Numerical results
• Test case: simple square domain with explicitly known covariance
15 July 2011
17 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Numerical results
• We can control the convergence rate of the series
• Problem loaded by a smooth boundary condition enforced with Nitsche‟s method
• Number of samples and truncation chosen as
15 July 2011
18 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Fast series - error
15 July 2011
19 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Fast series - workload
15 July 2011
20 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Moderate series - error
15 July 2011
21 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Moderate series - workload
15 July 2011
22 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Slow series - error
15 July 2011
23 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Slow series - workload
15 July 2011
24 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Implementation
• Monte Carlo methods are well-suited to parallelization
• Additional complications arise in the multi level framework – Load balancing less trivial
• Different memory and CPU requirements on sublevels
• Communication must be minimized
– Random series must be generated correctly
– Computing the series efficiently requires in-node parallelization
• Our combination
– C main routine running OpenMPI – MATLAB finite element solver
– OpenMP parallelization in mex-routine for K-L series computation
15 July 2011
25 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö
Finalised -
DBAB205397
Doc.ID: Revision: Status:
26
/ © Wärtsilä
Outlook and challenges
• Promising method with a wide variety of applications – Soil mechanics
– Corrosion modeling
• Computation of basis functions
– Explicit representation only in very simple geometries
• Should be computed numerically
• Estimation of convergence rate for eigenvalues
• Easy to implement with existing FE code
15 July 2011
26 The MLMC-FEM for a stochastic Brinkman Problem/ Juho Könnö