Tensor Sensitivity and long-range Coulomb interactions improve the accuracy and extensibility of Machine Learning Potentials

Large-scale atomistic simulations using fully ab initio forces remains an open problem due to the high costs of quantum chemistry solvers. Machine learning potentials offer an efficient alternative to direct quantum molecular dynamics. Many neural network architectures for machine learning potentials have relied only on short range interatomic distances. To overcome this simplification, we present the Hierarchically Interacting Particle Neural Network with Tensor Sensitivity, which processes geometric information via the full pairwise displacement vectors between nearby atoms, and Coulomb interaction extensions, which allow for physically constrained long-range interactions between atoms. We refer to this model as HIP-NN-TS-Coulomb. We trained on a data set of small water clusters, containing up to 14 molecules, and computed at the second order Møller–Plesset level of theory. When applied to the large scale simulation of liquid water, the model accurately reproduces properties such as radial distribution functions, diffusion constants, and density over a wide range of temperatures.