Atomistic neural network representations for chemical dynamics simulations of molecular, condensed phase, and interfacial systems: Efficiency, representability, and generalization

Abstract Machine learning techniques have been widely applied in many fields of chemistry, physics, biology, and materials science. One of the most fruitful applications is machine learning of the complicated multidimensional function of potential energy or related electronic properties from discrete quantum chemical data. In particular, substantial efforts have been dedicated to developing various atomistic neural network (AtNN) representations, which refer to a family of methods expressing the targeted physical quantity as a sum of atomic components represented by atomic NNs. This class of approaches not only fully preserves the physical symmetry of the system but also scales linearly with respect to the size of a system, enabling accurate and efficient chemical dynamics and spectroscopic simulations in complicated systems and even a number of variably sized systems across the phases. In this review, we discuss different strategies in developing highly efficient and representable AtNN potentials, and in generalizing these scalar AtNN models to learn vectorial and tensorial quantities with the correct rotational equivariance. We also review active learning algorithms to generate practical AtNN models and present selected examples of AtNN applications in gas‐surface systems to demonstrate their capabilities of accurately representing both molecular systems and condensed phase systems. We conclude this review by pointing out remaining challenges for the further development of more reliable, transferable, and scalable AtNN representations in more application scenarios. This article is categorized under: Data Science > Artificial Intelligence/Machine Learning Molecular and Statistical Mechanics > Molecular Interactions