GSM: GPU Accelerated Rare Events Sampling with Machine Learning Potentials

Abstract: Enhanced sampling has achieved considerable success in molecular dynamics (MD) simulations of rare events. Metadynamics (MetaD), owing to its excellent compatibility with MD engines, became one of the most popular enhanced sampling methods. With the boom of GPU computing and the advent of machine learning potentials (MLPs), high-accuracy, large-scale MD simulations have gradually become feasible. However, the corresponding GPU-based enhanced sampling tools have not yet been well adapted to this progress. To enable full-life-cycle GPU MetaD simulations, we propose the GPU Sampling MetaD (GSM) package. By leveraging MLPs, it is feasible to perform high-precision rare event sampling for systems comprising millions of atoms on a typical single GPU, which offers a potential solution to many size-dependent problems. By conducting sampling in several classical systems, the results sufficiently demonstrate the capability of this package to simulate diverse atomic systems, especially efficient in large scale systems.