Regensburg 2025 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 18: SYMD contributed

MM 18.7: Talk

Wednesday, March 19, 2025, 12:00–12:15, H23

Machine learning electrostatics: Open challenges from batteries to proteins — •Max Veit — Technische Universität Darmstadt, DE

Long-range interactions such as electrostatics have long been a concern in developing accurate, efficient machine learning potential energy surfaces (ML-PES). Such potentials have now become established as a powerful technique allowing simulations of complex structures and processes with unprecedented realism and accuracy. However, the most widespread and successful methods to date do not incorporate any interactions beyond a fixed cutoff range, typically a few coordination shells. First, we need to ask the question of how methods with such an obvious limitation can be so successful, even applied to systems where long-range electrostatic interactions are known to be relevant. Second, we need to ask what approach, among the many proposed over the last decade or so, is the most appropriate if we want to incorporate long-range interactions in an accurate, efficient, and physically appropriate way. We investigate these questions in the context of a technologically relevant, experimentally accessible test system: lithium-intercalated graphite or nearly-graphitic nanoporous carbon. We first discuss the characteristics of this system that make it uniquely suited to machine learning simulation, then turn to the difficulties involved in defining what exactly makes a ``good'' electrostatic model, or long-range model in general, in the context of machine learning potentials, and finally discuss the implications for other systems -- such as complex biomolecules -- just out of the current reach of ML-PES simulations.

Keywords: atomistic simulation; machine learning; potential energy surfaces; long-range interactions; electrostatics