(272f) Learning Many-Body Molecular Interactions from Machine Learning

The accurate representation of multidimensional potential energy surfaces is a necessary ingredient for realistic computer simulations of molecular systems. It has recently been shown that chemical and spectroscopic accuracy can be achieved with analytical potential energy functions (PEFs) rigorously derived from many-body expansions. In this contribution, we demonstrate the equivalence between permutationally invariant polynomials, neural networks, and Gaussian approximation potentials in representing lower-order many-body terms of molecular interactions in aqueous systems from the gas to the condensed phase. Besides demonstrating the synergy between high-quality electronic structure data and machine-learning techniques for developing transferable PEFs, our results show that machine learning provides a powerful tool for learning the underlying physics of many-body molecular interactions.