Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 22: Data-driven Materials Science: Big Data and Workflows III
MM 22.6: Talk
Wednesday, March 11, 2026, 11:45–12:00, SCH/A216
Data-efficient training of interatomic potentials using finite-temperature DFT structures — •Martin Schlipf1, Sudarshan Vijay1,2, and Georg Kresse1,3 — 1VASP Software GmbH, Berggasse 21/14, 1090 Vienna, Austria — 2Department of Chemical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai, Maharashtra 400076 India — 3Faculty of Physics and Center for Computational Materials Science, University of Vienna, Kolingasse 14-16, A-1090 Vienna, Austria
We successfully generated a database of 150,000 unique finite-temperature structures using VASP and a "one-shot" DFT method to systematically sample atomic environments across the periodic table. Despite the small size of our training set compared to the millions typically used for foundation models, our resulting interatomic potentials achieve a force prediction error of 72 meV/Å. This performance is of the same magnitude as current state-of-the-art foundation models when tested against the same high-quality dataset. This result demonstrates that focusing on data quality and chemical diversity at finite temperatures is as impactful as massive data quantity. Furthermore, we showcase the computational infrastructure that made it possible to integrate interatomic potentials into an ab-initio software and discuss necessary enhancements to electronic optimization methods to compute magnetic materials more reliably.
Keywords: DFT; VASP; Foundation Model; Machine Learning; Finite temperature