Quantitative assessment of long-range dipole–dipole interactions in the local ML representation

Determine quantitatively whether and to what extent long-range dipole–dipole interactions are captured by the unified differential learning model implemented in Allegro that uses a local representation of atomic environments for predicting electric enthalpy and derived response properties in extended systems.

Background

The proposed unified differentiable framework is implemented with an Allegro-based local, O(3)-equivariant neural network that enforces physical symmetries and conservation laws while enabling large-scale simulations. Because the model relies on strictly local atomic environment descriptions, the handling of inherently long-range electrostatic effects is not guaranteed by construction.

In the concluding discussion, the authors explicitly note that long-range dipole–dipole interactions may not be fully captured by the local model and acknowledge that the magnitude and significance of any missing contributions have not yet been quantified. They also argue that screening in extended homogeneous systems might mitigate such effects but emphasize that a quantitative investigation is still needed.