Berlin 2018 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 67: Methods in Computational Materials Modelling (methodological aspects, numerics)
MM 67.3: Talk
Thursday, March 15, 2018, 18:00–18:15, TC 006
Multilevel Adaptive Sparse Grids for parametric stochastic materials models — •Sandra Döpking1, Daniel Strobusch2, Christoph Scheurer2, and Sebastian Matera1 — 1Freie Universität Berlin — 2Technische Universität München
Many material models require sampling to obtain the desired model output, e.g. molecular dynamics or Monte Carlo sampling. Most of these models depend on a number of parameters which show a high variability. This can be an effect of uncertainties due to the limited accuracy of the underlying method. Or, the model might represent a class of materials with different material parameters. We present a Multilevel Adaptive Sparse Grid approach to explore the parameter space and to construct a surrogate of the often expensive original model. In this approach, the points in the parameter space are adaptively chosen which reduces the total number of costly model evaluations. Moreover, the multi-level structure of the sparse grids allows us to reduce the sampling accuracy and therefore the cpu-time spend for the model evaluation in each refinement step. We demonstrate the methodology for a first-principles kinetic Monte Carlo (1p-kMC) model for heterogeneous CO oxidation, where we address the impact of the uncertain reaction energetics derived from Density Functional Theory. We find that the multi-level approach reduces the computational cost significantly compared to non-adaptive, single level sparse grids -- without compromising the accuracy of the results. For this model, we observe that DFT uncertainty can have a tremendous impact on the simulation output and that 1p-kMC predictions have to be interpreted carefully.