Quantifying Trust in Interpretable Machine Learning for Materials Science

Machine learning (ML) models for materials properties can accelerate materials discovery. However, widely used ML models have limited interpretability and are trained on datasets that may not be representative of the materials space relevant to a specific application. To address these issues, we highlight recent advancements in the sure-independence screening and sparsifying operator (SISSO) [1], an interpretable, symbolic regression based ML method. Tailored uncertainty quantification approaches are developed for SISSO to improve both the reliability of predictions and the efficiency of materials discovery. These methods are further embedded into active learning workflows, enabling effective exploration of promising regions within the materials space [2]. We also introduce AI@FHI-aims [3], an online platform that integrates these developments with all-electron electronic structure simulations, designed for users with limited experience in this field.

(1) Ouyang, R.; Curtarolo, S.; Ahmetcik, E.; Scheffler, M.; Ghiringhelli, L. M. Physical Review Materials 2018, 2, 083802.

(2) Nair, A. S.; Foppa, L.; Scheffler, M. npj Computational Materials 2025, 11, 1–7.

(3) Abbott, J. W. et al. arXiv preprint arXiv:2505.00125 2025