Regensburg 2019 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 33: Topical session (Symposium MM): Big Data Analytics in Materials Science

MM 33.9: Vortrag

Donnerstag, 4. April 2019, 13:00–13:15, H43

Global sensitivity analysis and surrogate modeling for materials models with rapid local variations — Juan Lorenzi², Sandra Döpking¹, and •Sebastian Matera¹ — ¹Institut f. Mathematik, Freie Universität Berlin — ²Lehrstuhl F. Theoretische Chemie, Technische Universität München

Most material models depend on a number of input parameters which carry some uncertainty. Quantifying the impact of these errors on the model output is the purpose of global uncertainty and sensitivity analysis. This requires some kind of sampling of the parameter space and surrogate modeling has become a popular tool to lift the problem of repetitive, computationally expensive model evaluations. Surrogate modeling becomes challenging when the underlying model shows locally rapid variations, e.g. if a materials model exhibits a phase transition within the parameter domain. We present a modification of the classical Shepard interpolation, which has been designed for such problems. This approach employs a local, node specific distance metric instead of a global metric and uses error estimates for the superposition of different local linear models at a query point. We demonstrate the approach on the global sensitivity analysis of a stochastic model for CO oxidation, which has been parametrized using Density Functional Theory. We find that we can obtain reasonably accurate estimates of the sensitivity indices already at a modest number of evaluations of the original high-fidelity model.