Application of machine learning potentials in path integral molecular dynamics simulations

Thomsen, B.  ; Shiga, Motoyuki   

Path integral molecular dynamics (PIMD) and related methods, based on the Feynman path formulation of quantum mechanics, offer a direct way of modelling nuclear quantum effects in bulk phase materials. Each timestep of these methods does however require the evaluation of several electronic structure energies, gradients and stress vectors at the ab initio level. They are thus very computationally expensive to pursue at the ab initio level. Recently machine learned potentials (MLPs) have been suggested as a way to bring down the cost and allow long time dynamics to be studied with ab initio accuracy for PIMD methods. In this presentation we present the results of studying hydrogen diffusion in Palladium metal across several temperatures. We will also discuss our ongoing efforts to apply MLPs to PIMD studies of water and its isotopologues.