calorine: A Python package for constructing andsampling neuroevolution potential models
Eric Lindgren, J. Magnus Rahm, Erik Fransson, Fredrik Eriksson, Nicklas Österbacka, Zheyong Fan, Paul Erhart
Abstract
Molecular dynamics (MD) simulations are a key tool in computational chemistry, physics, and materials science, aiding the understanding of microscopic processes but also guiding the development of novel materials. A MD simulation requires a model for the interatomic interactions. To this end, one traditionally often uses empirical interatomic potentials or force fields, which are fast but inaccurate, or ab-initio methods based on electronic structure theory such as density functional theory, which are accurate but computationally very expensive (Müser et al., 2023). Machine-learned interatomic potentials (MLIPs) have in recent years emerged as an alternative to these approaches, combining the speed of heuristic force fields with the accuracy of ab-initio techniques (Unke et al., 2021). Neuroevolution potentials (NEPs), implemented in the GPUMD package, in particular, are a highly accurate and efficient class of MLIPs (Fan et al., 2021, 2022; Fan, 2022). NEP models have already been used to study a variety of properties in a range of materials, with recent examples including radiation damage in tungsten (Liu et al., 2023), phase transitions (Fransson, Wiktor, et al., 2023) and dynamics of halide perovskites (Fransson, Rosander, et al., 2023) as well as thermal transport in two-dimensional materials (Sha et al., 2023). Here, we present calorine, a Python package that simplifies the construction, analysis and use of NEP models via GPUMD.