(525k) Exploiting Machine Learned Interatomic Model Architecture for Enhanced Transferability, Efficiency, and Robustness

For decades, atomistic modelers interested in studying condensed phase reacting systems have been forced to choose between highly predictive yet computationally expensive first principles methods, and parametric “force field” approaches affording greater computational efficiency at the expense of accuracy and predictive power. But now, modelers have a third option in machine learned interatomic potentials (ML-IAP), which can yield models approximating quantum-based potential energy surfaces with computational expense that scales linearly with system size. These ML-IAPs have had a transformative effect by enabling “quantum accurate” simulations on previously inaccessible scales; however their (1) development for systems of more than a few atom types (especially those for which molecular chemistry is involved) necessitates large volumes of high expense (e.g., quantum) training data, and resulting models, and (2) they remain significantly more computationally intensive than classical “force field” models. In this presentation, we discuss recent advances in underlying form and fitting strategies for the ChIMES ML-IAP that aim to resolve these challenges by increasing model transferability and computational efficiency. Findings are discussed within the context of complex reacting materials.