Regensburg 2016 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 22: Topical session: Integrated computational materials engineering for design of new materials V

MM 22.2: Vortrag

Dienstag, 8. März 2016, 12:00–12:15, H39

Sampling parameter sensitivities for kinetic Monte Carlo models — Max J. Hofmann², Felix Engelmann³, and •Sebastian Matera¹ — ¹FU Berlin, Germany — ²U Stanford, USA — ³TU München

We present a three-step procedure for the sampling of local sensitivities (parameter gradients) from kinetic Monte Carlo models, which severely reduces the computational cost compared with standard numerical differentiation. In the first step, we utilize the Relative Entropy Method[1] for obtaining upper/lower bounds for the derivatives and discard those parameters with vanishing sensitivity. For the remaining, we sample the derivatives from a single trajectory employing an estimator based on Linear Response Theory. It turns out that some sensitivities can very well be sampled, while others show a higher variance. Only for the later, we perform a numerical differentiation using Coupled Finite Differences (CFD)[2]. We demonstrate the approach by revisiting the sensitivity analysis of the CO oxidation on RuO₂ (110)[3]. We find that CFD is only necessary for a small fraction of the sensitivities. [1] Y. Pantazis and M. A. Katsoulakis, J.Chem. Phys. 138 (2013): 054115; [2] D. F. Anderson, David, SIAM J. Num. Analysis 50 (2012): 2237-2258; [3] H. Meskine, S. Matera, M. Scheffler, K. Reuter, K., H. Metiu, (2009). Surf. Sci., 603 (2009): 1724-1730