Regensburg 2019 – wissenschaftliches Programm
MM 4.5: Vortrag
Montag, 1. April 2019, 11:15–11:30, H45
Uncertainty Quantification Framework for Force-Matched Classical Potentials — •Peter Brommer and Sarah Longbottom — University of Warwick, Coventry, UK
Classical effective potentials, where the energy of the system (and thus the forces on the atoms) only depend on the atomic positions, underpin a significant fraction of molecular dynamics (MD) simulations. Unfortunately, the bias, approximations and uncertainties incurred when representing the complex energy landscape by an effective potential or force field with a limited number of parameters are badly controlled, and the impact on quantities of interest (QoI) of an MD simulation is a priori unknown.
For potentials obtained by force matching, i.e. by adjusting their parameters to optimally reproduce forces, stresses and energies from first-principles calculations, we present an extension to the force-matching package potfit, where the uncertainty in the parameter set is encapsulated in an ensemble of potentials. This allows propagation of the uncertainty to QoI and makes uncertainty quantification accessible to users. We demonstrate this capability using three distinct potentials for Nickel and determine predictions and confidence intervals for lattice constants, elastic constants and thermal expansion.