Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 34: Poster: Machine Learning, Data Science, and Reservoir Computing
DY 34.12: Poster
Mittwoch, 20. März 2024, 15:00–18:00, Poster C
Long-Range Electrostatic Descriptors for Machine Learning Force Fields — •Carolin Faller1,2, Bernhard Schmiedmayer1, and Georg Kresse1,3 — 1Computational Materials Physics, University of Vienna, Austria — 2Vienna Doctoral School in Physics, University of Vienna, Austria — 3VASP Software GmbH, Vienna, Austria
We present flexible and physically meaningful descriptors for modeling long-range electrostatic interactions in machine learning force fields (MLFFs).
While local, atom-centered descriptors can accurately describe properties of several materials, they completely disregard long-range effects. For example long-range electrostatics are a crucial aspect of ionic materials, which makes it necessary to develop new techniques that account for them in order to ensure the predictive capability of MLFFs.
Our novel descriptors account for long-range interactions without resorting to a global description. They characterize the atomic density, similar to commonly used short-range methods. Periodic images of all atoms are accounted for by calculating the atomic density in reciprocal space.
This new long-range model is comparable to the long-distance equivariant (LODE) framework [1] for system with purely electrostatic interactions. Our model outperforms LODE in predicting energies and forces for real materials, where local approaches fall short.
[1] A. Grisafi, The Journal of Chemical Physics, 151, 204105 (2019).
Keywords: Machine Learned Potentials; Descriptors; Electrostatics; Long-Range; MLFF