Field-aware and charge-informed machine learning for predicting molecular and condensed-phase responses and vibrational spectra

Abstract External electric fields play a crucial role in chemistry, materials, and biology. However, accurately capturing external field effects remains a challenge for machine learning interatomic potentials (MLIPs). By revealing the limitations of conventional architectures in predicting field-dependent response properties, we introduce a physics-informed field-aware equivariant neural network framework that predicts response properties to external electric fields. External electric fields are incorporated through an equivariant embedding layer that concatenates field representations with atomic features, preserving rotational equivariance. The model leverages an energy-derivative approach to predict response properties in a single unified architecture. Furthermore, we embed a charge-equilibration electrostatic layer into the model to capture long-range electrostatic interactions. The close agreement between charged-based and energy derivative-based electric dipoles indicates that the model learns the correct underlying physics. The unified framework performs robustly for the prediction of response properties and simulation of IR and Raman spectra across molecular and periodic systems under different external field strengths. By incorporating physical principles into the model, we establish a general and transferable framework for modeling molecular and condensed-phase response under external electric fields, providing a robust routine for response properties prediction in external electric fields.