Leveraging the Potential of Machine-Learning Interatomic Potentials for QM/MM Simulations

Abstract

Machine-learning interatomic potentials (MLIPs) are increasingly used to replace computationally expensive quantum-mechanical (QM) calculations to obtain the energies and forces in ab initio or multiscale molecular dynamics (MD) simulations. While the computational cost of MLIPs lies between that of QM methods and classical force fields (molecular mechanics, MM), their accuracy is close to that of the chosen reference method (e.g. density functional theory, DFT) with sufficient training data. However, for large biological systems in solution, MLIPs are still too costly to perform long MD simulations, where the full system (i.e. including the solvent) is described by the MLIP. Instead, multiscale approaches analogous to QM/MM (i.e. ML/MM) offer a viable compromise between computational effort and accessible system size and time scales. In this review, we provide a brief overview of recent advances and current developments in this field.