Speaker
Description
Most atomistic machine-learning models rely on local atomic representations with finite spatial cutoffs and therefore struggle to capture long-range electrostatic effects. To address this limitation, we build on the fact that Born effective charges, defined as derivatives of the macroscopic polarization with respect to atomic displacements, provide a well defined description of the coupling between atomic structure and long-range electrostatic response. Unlike the macroscopic polarization itself, which is only defined modulo a polarization quantum in periodic systems, Born effective charges are unambiguous response quantities.
In our approach, local polarization dipoles are learned and constrained such that their derivatives reproduce the Born effective charges, and embedded in an explicit long-range electrostatic energy expression learned alongside the short-range part of the model, providing a consistent framework for coupling local atomic structure to long-range electrostatic interactions.We implement this within the Graph Atomic Cluster Expansion (GRACE), and demonstrate that the resulting model reduces prediction errors and mitigates finite-size effects, enabling models trained on small simulation cells to generalize to larger systems. Results are shown for representative material systems.