Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 18: New methods: Theory – Poster
O 18.1: Poster
Montag, 9. März 2026, 18:00–20:00, P2
Cost-Efficient Approach for Training MACE Potentials — •Antonia Gerstenberg1, 2, Thomas Bligaard2, and Andreas Lynge Vishart2 — 1currently: Fritz-Haber-Institut der MPG, Berlin, Germany — 2DTU, Department of Energy Conversion and Storage, Denmark
Machine learning interatomic potentials (MLIPs) such as MACE offer near DFT accuracy at a fraction of the computational cost. However, a limiting factor is the cost of generating high-quality training data. This project investigates the development of a systematic and cost-efficient approach to training robust MACE models. It is therefore explored to what extent models trained on the cheapest available structures (e.g., dimers and trimers) can extrapolate to larger structures such as nanoparticles and vice versa.
To test this, an active learning workflow was implemented, iteratively adding structures with increasing numbers of atoms. The results show that widely used training strategies, including standard active learning workflows, fail to reliably extrapolate to out-of-distribution structures. The results highlight the importance of structural diversity and weight restriction from the outset.
Keywords: MLIP; Machine Learning; MACE; Active Learning; out-of-distribution OOD