Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 9: Topical Session: Physics-driven Artificial Intelligence for Materials II

MM 9.2: Talk

Monday, March 9, 2026, 16:15–16:30, SCH/A251

Learning long-range interactions with equivariant charges — •Marcel F. Langer, Egor Rumiantsev, Tulga-Erdene Sodjargal, Michele Ceriotti, and Philip Loche — Laboratory of Computational Science and Modeling, École Polytechnique Fédérale de Lausanne, 1015 Lausanne, Switzerland

Machine-learning interatomic potentials trained on first-principles data have become key tools across computational physics, chemistry, and biology. Equivariant message-passing neural networks, including transformer variants, now deliver state-of-the-art accuracy, but their cutoff-based graphs restrict the treatment of long-range physics such as electrostatics, dispersion, and electron delocalisation. Existing long-range corrections based on inverse-power laws of distances capture only scalar interactions and cannot convey higher-order geometric information, limiting their applicability. To address this, we propose the use of equivariant (rather than scalar) charges to mediate long-range interactions and build a graph neural-network architecture, LOREM [1], around this equivariant message-passing scheme. The talk will outline the architecture, present results on several benchmark datasets, and discuss our work on universal long-range interatomic potentials.

[1] Egor Rumiantsev, Marcel F. Langer, Tulga-Erdene Sodjargal, Michele Ceriotti & Philip Loche, arXiv:2507.19382 (2025).