Machine Learning Interatomic Potentials for Molecular Simulations: Methods, Applications, Accuracy and Transferability

Abstract

The combination of the accuracy of quantum mechanical methods with the speed of classical force fields has made machine learning interatomic potentials (MLIPs) a transformative method in molecular and materials simulations. This review gives a complete picture of the basic principles, methodological advancements, applications, accuracy and transferability of the MLIPs. It first considers the theoretical background of potential energy surfaces, atomic interactions and necessary physical restrictions for a reliable interatomic modelling. The full development workflow is then summarized, from the generation of reference data to the generation of atomic descriptors, the training and validation of the model, hyperparameter optimization and the formulation of active learning strategies. The predictive power and computational complexity of major MLIP approaches, such as neural network potentials, Gaussian approximation potentials, moment tensor potentials, atomic cluster expansion models, graph neural networks, and equivariant potentials, are critically assessed. In addition, it considers approaches for the numerical evaluation of energies, forces, stresses, and structural property evaluations, and elucidates the need for careful validation, uncertainty quantification, and computational scalability. Transferability of MLIPs to various chemical compositions, phases, temperatures, defects, interfaces, and chemically reacting systems is also discussed, as is recent progress in foundation models and pretrained atomistic learning frameworks. Lastly, some challenges in the current state, such as limited reference datasets, long-range interactions, model interpretability, and benchmarking, are listed, and future research opportunities are discussed.