Explainable Data-driven Modeling of Adsorption Energy in Heterogeneous Catalysis

Explainability matters for scientific ML: The study demonstrates that combining machine learning with explainability techniques (specifically SHAP analysis and Symbolic Regression) can reveal why models make predictions, not just what they predict. This transparency is crucial for scientists to trust and validate AI-driven discoveries in catalysis research.

Adsorbate properties outweigh catalyst properties: The post-hoc analysis revealed that properties of the molecules being studied (adsorbates) have greater influence on predicting adsorption energy than the properties of the catalyst itself. This finding could redirect research priorities and resource allocation in catalyst design.

Interpretable models can match complex ones: The use of "shallow" machine learning techniques combined with explainability methods produced results consistent with symbolic regression, suggesting that simpler, more interpretable models may be sufficient for this application—potentially reducing computational costs and improving accessibility for researchers.

Key predictive features identified: The analysis pinpointed five critical features driving predictions: adsorbate electronegativity, number of adsorbate atoms, catalyst electronegativity, effective coordination number, and sum of atomic numbers. This specificity enables scientists to focus experimental efforts on the most impactful variables.

Bridging physics and data science: By integrating physics-based understanding with data-driven methods and explainability techniques, the research demonstrates a responsible AI approach that doesn't replace scientific knowledge but rather enhances and validates it.