Generalizing Abell-Tersoff bond-order potential with explicit high-order many-body correlations for robust extrapolation of potential energy surfaces

Abstract: Machine-learning interatomic potentials enable accurate and efficient atomistic simulations, but their reliability for out-of-distribution configurations far beyond the training domain remains a significant challenge. Here, to tackle this challenge, we introduce a semiparametric interatomic potential based on a generalization of the Abell--Tersoff bond-order potential, with a chemically informed functional form and explicit high-order many-body correlations. The model is trained and evaluated on various datasets, including silicon, carbon, water, and small molecules, achieving interpolation accuracy comparable to that of existing MLIP models while demonstrating better extrapolation to unseen configurations, including those under high-pressure and high-temperature conditions. These results demonstrate that physically motivated constraints incorporated into the functional form can improve the extrapolation behavior of interatomic potentials, providing insights into the inductive biases that can be used to control extrapolation in machine-learning interatomic potentials.