Regensburg 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 6: Data Analytics for Complex Systems (joint session DY/SOE)
SOE 6.5: Talk
Monday, September 5, 2022, 16:00–16:15, H18
Global sensitivity analysis of Monte Carlo models using Cramer-von Mises distance — •Sina Dortaj1,2 and Sebastian Matera1,2 — 1Fritz-Haber-Institut der Max-Planck-Gesellschaft, Faradayweg 4-6, 14195 Berlin, Germany — 2Institute for Mathematics, Freie Universität Berlin, Arnimallee 6, 14195 Berlin, Germany
Typically, the parameters entering a physical simulation model carry some kind of uncertainty, e.g. due to the intrinsic approximations in a higher fidelity theory from which they have been obtained. Global sensitivity analysis (GSA) targets quantifying which parameters uncertainties impact the accuracy of the simulation results, e.g. to identify which parameters need to be determined more accurately.
We present a GSA approach on basis of the Cramers-von Mises distance. Unlike prevalent approaches it combines the following properties: i) it is equally suited for deterministic as well as stochastic model outputs, ii) it is free of gradients, and iii) it can be estimated from any suitable numerical quadrature (NQ) without further numerical tricks. Using Quasi-Monte Carlo for NQ and prototypical first-principles kinetic Monte Carlo models (kMC), we examine the performance of the approach. We find that the approach typically converges in a modest number of NQ points. Furthermore, it is robust against even extreme relative noise. All these properties make the method particularly suited for expensive (kinetic) Monte Carlo models, because we can reduce the number of simulations as well as the target variance of each of these.