Accurate learning of long-range interatomic potentials by coupling Cartesian atomic cluster expansion and sum-of-Gaussians neural networks

Accurately capturing long-range interactions is critical for molecular dynamics simulations based on machine learning interatomic potentials. We recently proposed the sum-of-Gaussians neural network (SOG-Net), which learns long-range energy contributions directly from energy and force data such that the long-range tail of different decay rates can be well fitted. In this work, we incorporate the SOG-Net with a short-range descriptor of the Cartesian atomic cluster expansion, resulting in the CACE-SOG model, to show that the SOG-Net is a general module that can be coupled with different short-range descriptors. We also study new technical developments in the SOG-Net, including improved extrapolation accuracy, handling of different charge states, and faster convergence. We evaluated the CACE-SOG model across a diverse set of systems, including molecular dimers, aqueous salt solutions, charged ionic clusters, and liquid-vapor and Pt(111) interfacial water systems, and compared it with the CACE-based latent Ewald summation and the CACE-only methods. These results demonstrate that the SOG-Net is promising for accurately learning long-range interatomic interactions.