Property-Defining Feature Extraction in Porous Crystalline Materials Using Attention-Guided Cooperative Game Theory

Porous crystalline materials assembled from molecular building blocks offer vast potential across applications such as gas storage, separation, and catalysis, yet their structural tunability creates an expansive design space that demands the systematic identification of their property-defining features. To help researchers identify these features rapidly and inform next generation design, we introduce the Explainable Crystal Transformer (ExCT); an attention-guided cooperative game-theoretic workflow for extracting property-defining structural features in porous crystalline materials by combining attention-based representation learning, Shapley-style attribution, and local surrogate-fidelity screening. Designed for broader use across reticular frameworks, ExCT is demonstrated here with metal–organic frameworks (MOFs) as a representative case study of H2, CH4, and CO2 adsorption: cooperative game-theoretic attributions highlight chemically plausible local regions in selected hypothetical MOFs and qualitatively correspond to reported adsorption environments in experimentally studied cases, whereas attention identifies broader property-associated regions. Strict held-out surrogate-fidelity screening retains only a subset of high-fidelity explanations, more frequently for H2 than CH4 and less frequently for CO2.