Acceleration of Simulations of Complex Compounds by Dimensionality Reduction of Atomic Cluster Expansion Features

Machine learning interatomic potentials have shown promising accuracy and efficiency with equivariant features. Among those, atomic cluster expansion [1] is proposed recently as a systematic method to expand body orders of the features. However, the dimension of atomic cluster expansion features grows quickly with the number of chemical elements, which limits its performance for complex materials. In this work, we investigate embedding strategies for dealing with multiple chemical elements and basis functions of atomic cluster expansion. Our theoretical and numerical results illustrate that the embedding method is able to reduce the dimension of the features with almost no loss of accuracy, yielding significant acceleration of machine learning molecular dynamics of complex materials.<br/><br/>[1] Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials, 2019