(197s) Learning Analytical Bond-Order Parameters for Extrapolatable Neural Network Interatomic Potential of Multicomponent Materials

Parametrization of neural network interatomic potentials has been widely studied to enable accelerated atomistic simulations at quantum chemical accuracy. The method proposed by Behler and Parrinello uses atom-centered symmetric functions (ACSF) to encode both geometric and chemical information, predicting atomic energy with high-dimensional neural networks. While this approach is powerful for fitting interatomic potentials to ab-initio calculations, limited extrapolation capability remains a significant challenge since the computational cost of data acquisition is notoriously high.

To address this issue, we introduce a regularization layer based on the analytical bond-order potential (BOP) model, imposing physical inductive bias to the model. Pairwise bond-order parameters are represented as functions of the local atomic environment using neural networks and are utilized in energy prediction with BOP. Our model demonstrates remarkable extrapolation capabilities for potential energy across unseen molecular geometries, which is unprecedented in conventional Behler-Parrinello neural network models.

Moreover, our model offers a physically meaningful interpretation of the learned potential energy by decomposing it into pairwise interactions. We validate the energy decomposition by comparing it with the quantum-chemically derived intrinsic bond strength index (IBSI), which shows excellent agreement. We demonstrate the pairwise interaction analysis to the path of the S_N2 reaction of methyl halides. It accurately captures the evolution of interatomic interactions throughout the reaction, showing the applicability of model on the study of chemically reactive systems.