Bayesian uncertainty quantification and propagation in molecular dynamics simulations

Abstract

A comprehensive Bayesian probabilistic framework is developed for quantifying and calibrating the uncertainties in the parameters of the models (e.g. force-field potentials) involved in molecular dynamics (MD) simulations based on experimental data, as well as propagating these uncertainties for the prediction of uncertainties for other quantities of interest. Conceptual issues as well as computational High Performance Computing (HPC) issues encountered in Bayesian uncertainty quantification and propagation (UQ+P) in MD simulations are addressed. Stochastic simulation algorithms (e.g. MCMC) requiring multiple MD runs are demonstrated to be the only viable alternatives for UQ+P in MD simulations. A transitional MCMC approach for fast trace of the important uncertainty region in the model parameter space is demonstrated to have the best performance among various stochastic simulation algorithms for populating the posterior probability distribution of the model parameters. Parallel computing algorithms and surrogate models are well suited to be integrated with TMCMC algorithm. Parallel computing algorithms are proposed to efficiently handle the large number of MD model runs and distribute the computations in available GPUs and multi-core CPUs in heterogeneous architecture clusters. They exploit the large number of independent chains created and simultaneously run in the TMCMC stages. Surrogate models are adopted to drastically reduce the number of MD model runs achieving additional substantial computational savings.The effectiveness and computational efficiency of the proposed Bayesian framework is demonstrated using MD simulations for molecules of argon.