Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 19: Poster Session
MM 19.10: Poster
Dienstag, 10. März 2026, 18:00–20:00, P5
Transfer Learning Pipeline for GRACE Foundation Models for Complex Materials — •Christian L. Ritterhoff1, Yury Lysogorskiy2, Anton Bochkarev2, Bernd Meyer1, and Ralf Drautz2 — 1Interdisciplinary Center for Molecular Materials (ICMM) and Computer Chemistry Center (CCC), FAU Erlangen-Nürnberg — 2Interdisciplinary Centre for Advanced Materials Simulation (ICAMS), Ruhr-Universität Bochum (RUB)
Foundation models of machine-learned interatomic potentials (MLIPs) offer overall good accuracy for a wide range of configurational and chemical spaces. However, achieving high fidelity and high efficiency for dedicated applications often requires further refinement of the model. This is commonly achieved by fine-tuning the foundation model on minimal specialized training data followed by distillation of the acquired knowledge into smaller models. We benchmark different strategies for the example of carbon with its demanding diverse chemistry and structure by fine-tuning the GRACE-1L-OMAT foundation model using the dataset by Qamar et al. [1]. Data efficiency is analyzed by training only on randomly chosen subsets of the complete dataset. Advantages of fine-tuning are demonstrated by comparing the obtained models against their randomly initialized counterparts. Finally, the resulting high-fidelity GRACE potential is used for knowledge distillation into a fast, local ACE model. This work validates the transfer learning and distillation paradigm as a robust and efficient pathway for creating deployable potentials for complex materials.
[1] M. Qamar, et al., J. Chem. Theo. Comput. 19 (2023)
5151–5167
Keywords: machine-learned interatomic potentials; foundation models; fine tuning; workflows