Uncertainty Quantification in Molecular Dynamics Simulations


Wednesday, 21 May, 2014 - 11:30

Speaker: Francesco Rizzi (Duke University)

Room: SISSA - Santorio A - room 133

Molecular Dynamics (MD) simulations provide a suitable tool to explore the properties of a system at the atomic level which, in general, are difficult and expensive to investigate experimentally. The main weakness of MD is that its predictive reliability depends on the accuracy with which the MD potential function can model the atomic interactions occurring in the real system of interest. Consequently, defining the potential is the most delicate stage of an MD simulation. This is typically done in a deterministic setting, namely by choosing specific values for the parameters of the MD potential. Literature, however, shows that for most MD systems, these parameters are characterized by broad uncertainties. Uncertainty quantification (UQ) can thus play a key role for quantifying these uncertainties, and properly characterize the predictive accuracy. This talk shows a possible approach for applying UQ methods to MD simulations. Two fundamental, distinct sources of uncertainty are investigated, namely parametric uncertainty and intrinsic noise. Intrinsic noise is inherently present in the MD setting, due to fluctuations originating from thermal effects. Parametric uncertainty, on the contrary, is introduced in the form of uncertain potential parameters, geometry, and/or boundary conditions. We illustrate the use of a probabilistic (Bayesian) approach to infer atomistic parameters for MD simulations of pure water using data of selected macroscale observables. Using Polynomial Chaos (PC) expansions and Bayesian inference, we develop a framework that enables us to describe the impact of parametric uncertainty on the MD predictions and, at the same time, properly quantify the effect of the intrinsic noise.