Evolutionary machine learning of physics-based force fields in high-dimensional parameter-space

Abstract

This work presents the Alexandria Chemistry Toolkit (ACT), an open-source software for machine learning of physics-based force fields (FFs) from scratch, based on user-specified potential functions. In this approach, a set of FF parameters for molecular simulation is described as a chromosome consisting of atom and bond genes. The accuracy of a FF, that is how well quantum chemical training data are reproduced, determines the fitness of the chromosome. The ACT implements a hierarchical parallel scheme that iterates between a genetic algorithm and Monte-Carlo steps for global and local search, to find “genomes” with high fitness. As a sample application, genome evolution is performed to create physical models that allow the prediction of properties of organic molecules in the gas and liquid phases. Evaluation of the prediction accuracy of different models showcases how force field science can contribute to systematically improve prediction accuracy of physicochemical observables.