Continuum Mechanics Group

Predictive computational modeling

n The focus of the Continuum Mechanics Group in 2015 was the development of novel models, methodologies and computational tools for quantifying uncertainties and their effect in the simulation of engineering and physics systems. Our work has been directed towards three fronts:

a) the calibration and validation of computational models using experimental data, b) uncertainty propagation in multiscale systems, c) Design/control/ optimization of complex systems under uncertainty.

Prof. Dr. Phaedon-Stelios

Koutsourelakis

Contact

www.contmech.mw.tum.de Phone +49.89.289.16690

A highlight was the organization of the international Symposium on ‘Big Data and Predictive Computational Modeling’ that took place in TUM-IAS during May 18-21

2015. The symposium was sponsored by the European Office of U.S. Aerospace Research Development (EOARD), the TUM-IAS and the Department of Mecha- nical Engineering. It included plenary and

keynote talks from pre-eminent scientists

in applied mathematics, computational physics/chemistry, computer science and engineering.

The event received considerable interest from various communities and an article on it appeared in Volume 48, Number 6

of SIAM news. Videos and slides from the symposium can be found at: http://www. tum-ias.de/bigdata2015/program.html

Excerpt from SIAM news article (Volume 48, Number 6) on ‘Big Data and Predictive Computatio- nal Modeling’

Nonlinear Inverse Problems with Applications in Medical Diagnostics

This project is concerned with the numer- ical solution of model-based, Bayesian inverse problems. We are particularly interested in cases where the cost of each likelihood evaluation (forward-model call)

is expensive and the number of unknown (latent) variables is high. This is the setting of many problems in computational phy- sics where forward models with nonlinear PDEs are used and the parameters to

be calibrated involve spatio-temporarily varying coefficients, which upon discre- tization, give rise to a high-dimensional vector of unknowns.

Our motivating application is biome- chanics where several studies have shown that the identification of material parameters from deformation data can lead to earlier and more accurate diag-


nosis of various pathologies. This project develops Bayesian computational tools that can lead not only to point-estimates of the material properties, but also to a rigorous quantification of the confidence

in those estimates. A significant role in this

effort is played by novel dimensionality reduction techniques which can identify a sparse set of features. One consequence of the ill-posedness of inverse problems

is multi-modality i.e. the possibility of multiple, distinct solutions that fit the data comparably well. We developed adaptive strategies that are capable of accurately capturing the presence of multiple modes.

Coarse-Graining in Equilibrium Statistical Mechanics

(a) Ground truth (reference) (b) Multimodal posterior mean. Points on the boundary have non-zero probability of belonging to healthy

tissue (blue) or tumor (red)

Identification of material para- meters using nonlinear, incom- pressible elasticity models.

Coarse-grained (CG) models provide a

computationally efficient means to study large numbers of atoms over extended spatio-temporal scales. Existing strategies rely on mapping degrees of freedom from micro- to macro-scale, whereas properties are commonly determined as certain

point-estimates in the macro-scale. Various methodologies that have been proposed do not account for the informa- tion loss which unavoidably takes place during coarse-graining. The joint project

‘Predictive Materials Modeling’ with Hans Fischer Senior Fellow Prof. N. Zabaras (Director of Warwick Centre for Predictive Modeling, University of Warwick) addres- ses the formulation and development of

(a) Coarse-graining of water molecules

an enhanced, predictive multiscale frame- work. Our reformulation of coarse-graining follows a data-driven, Bayesian paradigm based on generative probabilistic models. It builds upon: a) a coarse macroscopic description, and b) a probabilistic lifting operator, reconstructing micro configu- rations from coarse states. This affords a more flexible definition of CG variables,

the rigorous quantification of uncertainties

associated with using limited training data and various degrees of information loss. Furthermore, it is capable of iden- tifying sparse representations for the CG potential which reveal qualitative, physical features of the CG model.

(b) Proposed coarse-graining scheme

Reduced-order Modeling for Uncertainty Quantification

(a) True (reference) solution manifold in three-dimensional phase space.

Reduced-basis construction for Kraichnan-Orszag nonlinear ODE with parameters/inputs being the initial conditions and time.


(b) Mixture of local, reduced-basis identified from data. Each color corresponds to a different reduced-basis set associated with a particular subdomain of the input space.

As the physical problems become more complex and the mathematical models more involved, current computational methods prove increasingly inefficient,


especially in tasks requiring numerous solutions as in uncertainty quantification. In the latter case, the difficulty is further amplified by the large number of input parameters (random variables). This project employing advanced, statistical learning tools in order to develop probabi-

listic reduced-basis models. These consist of mixtures of reduced-bases which can

be used to project the governing equa- tions in order to obtain descriptions of significantly lower dimension. The tools developed can learn reduced-bases sets from a limited number of full-scale runs and can simultaneously partition the input

and output space so that different response regimes can be computed as well as the most important/sensitive inputs can be

identified.

Topology optimization problem. Deterministic (left) vs. Stochastic (right) solutions. The stochastic version accounts for random vari- ability in the material properties


Design in the Presence of Uncertainty

(a) Deterministic solution (b) Stochastic solution (mean)

This project is concerned with a lesser- studied problem in the context of

model-based, uncertainty quantification

(UQ), that of optimization/design/control under uncertainty. The solution of such problems is hindered not only by the

usual difficulties encountered in UQ tasks

(e.g. the high computational cost of each forward simulation, the large number of random variables) but also by the need

to solve a nonlinear optimization prob- lem involving large numbers of design variables and potentially constraints. We propose a framework that is suitable for a class of such problems and is based on the idea of recasting them as proba- bilistic inference tasks. To that end, we


developed a variational Bayesian (VB) formulation and an iterative VB expecta- tion maximization scheme that is capable of identifying a local maximum as well as a low-dimensional set of directions in the design space, along which, the objective exhibits the largest sensitivity. In cases considered the cost of the computations

in terms of calls to the forward model was of the order of 100 or less. The accuracy of the approximations provided was

assessed by information-theoretic metrics.

Research Focus

n Uncertainty quantification

n Random media

n Multiscale formulations

n Bayesian inverse problems

n Design/optimization under uncertainty

Competence

n Computer simulation

n Mathematical modeling of stochastic systems

Infrastructure

n 256core HPC

Courses

B.Sc.

n Uncertainty Quantification in Mechani- cal Engineering (SS)

n Modeling in Structural Mechanics (WS)

Master

n Atomistic Modeling of Materials (WS)

n Bayesian Strategies for Inverse Prob- lems (SS)

n Journal Club Uncertainty Quantification

(WS-SS)

MSE

n Continuum Mechanics (WS)

n Uncertainty Modeling in Engineering

(SS) (Top Teaching Trophy 2014, 2015)


Management

Prof. Phaedon-Stelios Koutsourelakis, Ph.D., Director

Administrative Staff

Cigdem Filker

Research Scientists Dipl.-Ing. Isabell Franck Markus Schoeberl, M.Sc.

Constantin Grigo, M.Sc. (Physics) Maximilian Koschade, B.Sc. Mariella Kast, B.Sc.

Lukas Koestler, B.Sc. Maximilian Soepper, B.Sc.

Publications 2014-15

n I. Franck, P.S. Koutsourelakis. Sparse Variational Bayesian Approximations for Nonlinear Inverse Problems: Applications in Nonlinear Elastography, Computer Methods in Applied Mechanics and Engineering, in press 2015 (to appear Volume 299,

1 February 2016, pp. 215-244).

n P.S. Koutsourelakis. Variational Bayesian Strategies for High-Dimensional, Stochastic Design Problems. Journal of Computational Physics, accepted 2015 (in press).

n I. Franck, P.S. Koutsourelakis. Multimodal,

high-dimensional, model-based, Bayesian inverse problems with applications in biomechanics, in preparation arxiv

n I. Franck, Uncertainty Quantification for Nonlinear

Inverse Problems in High DDimensions. 2nd poster award, 3rd ECCOMAS International Young Investi- gators Conference, Aachen, Germany (poster).

n M. Schoeberl. A Bayesian Approach to Coarse-

Graining. Predictive Multiscale Materials Modelling. Isaac Newton Institute, Cambridge. 2015 (poster).


n M. Schoeberl, N. Zabaras, P.S. Koutsourelakis.

Predictive Coarse-Graining. Predictive Multiscale Materials Modelling. Isaac Newton Institute, Cambridge. 2015 (presentation).

n I.Franck, P.S. Koutsourelakis. Variational Bayesian

Formulations for High-Dimensional Inverse Problems. SIAM CSE 2015, Salt Lake USA. 2015 (presentation).

n P.S. Koutsourelakis. Variational Bayesian Strategies

for High-Dimensional, Stochastic Design Problems. Big Data and Predictive Computational Modeling, May 2015, Germany (presentation).

n M. Schoeberl, N. Zabaras, P.S. Koutsourelakis. Pre-

dictive Coarse-Graining. To be submitted to Journal of Computational Physics 2015 (in preparation).