Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 5: Topical Session: Physics-driven Artificial Intelligence for Materials I
MM 5.3: Talk
Monday, March 9, 2026, 11:00–11:15, SCH/A251
Interpretable Bayesian Optimization for Autonomous Materials Discovery — •Akhil S. Nair1,2, Lucas Foppa1, and Matthias Scheffler1 — 1The NOMAD Laboratory at the FHI of the Max Planck Society, Berlin, Germany — 2Institut für Chemie und Biochemie, Freie Universität Berlin, Germany
Bayesian Optimization (BO) can accelerate materials discovery by exploring complex design spaces using surrogate models and acquisition functions [1]. Its efficiency, however, relies on identifying a small set of key parameters or features that are potentially correlated with the target property. Existing feature-selection methods often fall short, as they struggle to capture nonlinearities and interactions among features [2], limiting BO’s performance in high-dimensional spaces. To overcome this challenge, we introduce the Sparse Adaptive Representation-based Bayesian Optimization (SARBO) framework, which integrates BO with the Sure-Independence Screening and Sparsifying Operator (SISSO) method [3]. By capturing the non-linear interactions, SARBO identifies the most relevant features and adaptively updates their selection during the BO cycles, ensuring the optimization is continuously guided by the features that matter most. We demonstrate SARBO’s effectiveness through the simulated discovery of single-atom alloy catalysts for CO2 activation.
[1] Y. Tian, et al., npj Comput. Mater. 11, 209 (2025)
[2] M. R-. Kochi et al., Chem. Sci. 16, 5464 (2025)
[3] R. Ouyang et al., Physical Review. M 2, 8 (2018)
Keywords: Materials Discovery; Bayesian Optimizaion; SISSO; Feature Selection