Regensburg 2025 – wissenschaftliches Programm

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

MA: Fachverband Magnetismus

MA 35: PhD Focus Session: Using Artificial Intelligence Tools in Magnetism

MA 35.1: Hauptvortrag

Donnerstag, 20. März 2025, 09:35–10:05, H20

Artificial Intelligence for Materials Science: Critical Importance of Rare Events, Active Learning, and Uncertainties — •Matthias Scheffler — The NOMAD Laboratory at the Fritz Haber Institute of the Max Planck Society, 14195 Berlin, DE

Materials properties are often governed by an intricate interplay of many processes. As a consequence, the description in terms of meaningful analytical equations is typically inappropriate, and we are promoting the concept of 'materials genes'. These are elemental materials features that 'correlate' with the materials property of interest. Thus, they address the full intricacy and describe (in a statistical sense) the material's property and function.[1]

AI and machine learning (ML) exhibit diminished reliability when entering uncharted data regions. When the training data are representative of the full population (or iid), extrapolation may work. However, for materials this requirement is hardly fulfilled. Still, materials scientists are searching for 'statistically exceptional' situations, and properties are often triggered by 'rare events' that are not or not well covered by the available data, or smoothed out by the ML regularization. This all implies caution when applying ML. In my talk I will explain these issues and routes toward solutions. Key issues are the 'range of applicability' of ML models, the awful overconfidence of prediction uncertainties, and the needs for active learning.

(**) In collab. with Lucas Foppa, Kisung Kang, and Akhil S. Nair.

1) Scheffler M AI guided workflows for screening the materials space. Coshare Science 02, 02 (2024); https://doi.org/10.61109/cs.202403.129

Keywords: artificial intelligence; machine learning; materials science; chemical physics; uncertainty quantification