Unveiling the Core of Materials Properties via SISSO and Sensitivity Analysis: Use-case Demonstration for Perovskites

Interpretable AI can help reveal the physical principles governing intricate material properties and functions. In particular, the sure-independence screening and sparsifying operator (SISSO) symbolic-regression approach identifies analytical expressions correlating a target materials performance to a small set of physical descriptive parameters, termed materials genes, selected from a vast pool of primary features. However, the identified genes influence the SISSO models to different degrees. Here, we use the gradient-based partial-effect sensitivity analysis to pinpoint the most influential genes, thus enhancing SISSO's interpretability and enabling deeper physical insights. This analysis also highlights that different combinations of genes can yield equally accurate descriptions of the correlation. The approach is demonstrated for the bulk properties of perovskites.