Machine learning-assisted descriptor screening of high-entropy alloy-based catalysts for CO2 methanation

Carbon dioxide hydrogenation provides a route for converting CO 2 into methane (CH 4 ), but the large compositional space of high-entropy alloy (HEA)-based catalysts complicates empirical catalyst selection. In this study a literature-derived dataset of 251 observations was analysed using machine learning to evaluate composition-derived atomic and electronic descriptors together with surface, synthesis, and reaction variables. XGBR achieved random-split test R 2 values of 0.95 for CO 2 conversion and 0.91 for CH 4 selectivity, whereas intrinsic-descriptor-only models achieved R 2 values of 0.37 and 0.69, respectively, showing that full-model performance arises from combined catalyst and process information. SHAP and partial-dependence analyses identified reaction temperature and composition-derived descriptors including weighted Pauling electronegativity (EN), weighted atomic radius (R), and ΔR as influential model features. Weighted Pauling EN above approximately 1.2, smaller weighted atomic radii (R), and low ΔR were associated with favourable model predictions within the present dataset. These relationships are interpreted as empirical screening trends rather than direct mechanistic thresholds. NSGA-II and fuzzy logic analysis defined descriptor space screening regions, and more than 205 million five-element compositions were enumerated, of which 4,811,467, or 2.35%, satisfied the property window criteria. Experimental assessment showed lower apparent activation energies than Ni/Al 2 O 3 for several HEA-based formulations, with AlCdMgCoNi achieving 45 kJ mol −1 compared with 75 kJ mol −1 for the reference. However, the synthesized catalysts were multiphase, CH 4 selectivity remained below 60% at atmospheric pressure, and selectivity predictions transferred poorly to the experimental systems. These limitations indicate that surface-sensitive descriptors and descriptors capturing the local chemical environment are required for selectivity-focused catalyst design. Therefore, the framework is supported primarily as an activity-oriented screening and property-space narrowing approach.