Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 9: Poster

MM 9.60: Poster

Montag, 17. März 2025, 18:30–20:30, P1

Fine-tuning of machine learning interatomic potential for the prediction of phonon properties — •Jonas Grandel, Philipp Benner, and Janine George — Bundesanstalt für Materialforschung und Prüfung, Berlin

Accurate phonon predictions are critical for assessing material stability and thermal behavior, but traditional approaches based on density functional theory (DFT) are computationally expensive, motivating the need for accelerated alternatives. In this work, we investigate the performance of the machine learning interatomic potential MACE-MP-0 for predicting harmonic phonons and thermal properties. The focus is on fine-tuning MACE-MP-0 using various sets of rattled structures and different hyperparameter to identify the most effective strategy for improving model accuracy. We want to develop a general fine-tuning workflow based on the foundational model that can be used to fast and accurately generate phonons to predict both stability and thermal properties. For this purpose, a benchmark dataset was constructed using DFT consisting of a broad range of different crystal systems and mainly of phase change materials and thermoelectric materials. Each fine-tuned model targets one specific material, allowing to improve each material individually. The results demonstrate significant improvements in the prediction of phonon band structures, with a root mean square error (RMSE) reduced from 0.6 THz for the original MACE-MP-0 model to 0.3 THz for the fine-tuned models. In addition, performance in terms of computational speed was improved by up to a factor of 10 compared to traditional DFT-based phonon calculations.

Keywords: Machine Learning; Phonons; Interatomic Potential; Fine-tuning