Dresden 2026 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 22: Data-driven Materials Science: Big Data and Workflows III

MM 22.8: Vortrag

Mittwoch, 11. März 2026, 12:15–12:30, SCH/A216

Benchmarking the MACE Foundation Model for Solid-State Ion Conductors — •Takeru Miyagawa, Yufeng Xu, Levon Satzger, Waldemar Kaiser, and David A. Egger — Physics Department, TUM School of Natural Sciences, Technical University of Munich, 85748 Garching, Germany

Recent progress in foundation model machine learning potentials (MLPs) has demonstrated promising transferability and accuracy across diverse material classes [1, 2]. Instead of being trained from scratch for each new system, these large pretrained models aim to provide broadly accurate force and energy predictions that can be refined for new chemistries with comparatively small datasets. This offers a complementary route to traditional system-specific MLPs and may reduce the cost of studying complex ionic materials.

Here, we benchmark the MACE foundation model [2] on representative solid-state ion conductors (SSICs) through direct comparison with first-principles calculations. We assess its accuracy for phonons and vibrational properties, characterize temperature-driven structural and phase transitions, and analyze ion transport across different phases. We then explore data-efficient DFT-based fine-tuning strategies to improve the foundation model's accuracy for SSICs and clarify the limits and strengths of pretrained representations in the context of ionic transport. References [1] Batatia, I. et al., Adv. Neural Inf. Process. Syst. 35, 11423-11436, 2022, [2] Batatia, I. et al., J. Chem. Phys. 163, 184110, 2025

Keywords: Foundation models; Machine learning potentials; Lattice dynamics; Ion transport