Berlin 2024 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 43: Data Driven Material Science: Big Data and Workflows V

MM 43.4: Talk

Wednesday, March 20, 2024, 16:30–16:45, C 243

A many-body framework for long-range interactions in atomistic machine learning — •Kevin Kazuki Huguenin-Dumittan, Philip Robin Loche, and Michele Ceriotti — EPFL, Lausanne, Switzerland

Many properties of matter, from chemical bonds to the band structure in solids, arise from the quantum nature of electrons. Accurate atomistic modeling of such systems thus requires the use of quantum approaches, which can computationally become prohibitively expensive compared to their classical counterparts. Machine learning based approaches have seen a surge in interest over the past decade to bridge the gap between the two worlds by providing near-quantum accuracy at a cost scaling similarly to classical methods. One systematic source of error in most such models is the assumption of locality, which neglects important long-range interactions ranging from electrostatics to more complex effective interactions between macromolecules. In this talk, we present a long-range framework that is a significant improvement over the previously introduced LODE descriptors, which (1) scales effectively linearly with system size, (2) captures a broad family of long-range interactions beyond the Coulomb potential, (3) can systematically generate many-body features beyond just pair potentials and (4) smoothly and fully integrates with preexisting machine learning frameworks with little extra effort including the computation of gradients or equivariant models for tensorial properties.

Keywords: machine learning; many-body; long-range; equivariant; electrostatics