Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 3: Data-driven Materials Science: Big Data and Worksflows
MM 3.6: Vortrag
Montag, 17. März 2025, 11:45–12:00, H10
Accelerating Materials Exploration with Active Machine Learning: Integrating SISSO with FHI-aims — Yi Yao1,2, Lucas Foppa1, Akhil Sugathan Nair1, Andrei Sobolev1,2, •Konstantin Lion1,2, Sebastian Kokott1,2, and Matthias Scheffler1 — 1NOMAD Laboratory at the Fritz Haber Institute of the Max Planck Society, Berlin, Germany — 2Molecular Simulations from First Principles e.V., Berlin, Germany
We present a user-friendly web application for active learning-based materials exploration with the goal of broadening the usability of AI tools. The platform integrates the SISSO (Sure Independence Screening and Sparsifying Operator) method [J. Chem. Phys. 159, 114110 (2023)] with FHI-aims software [Comp. Phys. Commun. 180, 2105 (2009)] to provide interpretable modeling and reliable property predictions. SISSO dynamically updates models during the exploration process, while FHI-aims ensures accurate all-electron density functional theory (DFT)-based calculations. The property prediction workflow is managed using the atomate2 library, providing many "standard" DFT workflows and efficient utilization of compute resources ranging from local machines to cloud infrastructures. By leveraging SISSO-based uncertainty prediction, the application implements active learning to efficiently identify materials with desirable target properties. Two case studies, the exploration of the bulk modulus in perovskites and the prediction of stable oxides under harsh conditions, demonstrate the platform's ability to accelerate materials discovery.
Keywords: Machine Learning; Materials Discovery; SISSO; Density Functional Theory