High-Throughput Screening and Interpretable Machine Learning for Rational Design of Bimetallic Catalysts for Methane Activation.

Methane's efficient catalytic removal is vital for sustainable development. Bimetallic catalysts, though promising for methane activation, pose a design challenge due to their complex compositional space. This work introduces an integrated framework that combines high-throughput density functional theory (DFT) and interpretable machine learning to accelerate the rational design of catalysts. Computational screening of face-centered-cubic (FCC) bimetallic catalyst surfaces identifies the bond cleavage energies of the first and the second C─H bonds and methyl adsorption energy as a key descriptor governing successive C─H activation and the shift in the rate-determining step (RDS). Through the synergistic interaction of these descriptors, machine learning models can be constructed more effectively, leading to the discovery of a bimetallic catalyst for consecutive C─H bond cleavages that outperforms conventional natural gas engine aftertreatment systems. Based on the computationally derived DFT dataset, four machine learning models were trained using a particle swarm optimisation (PSO) algorithm, from which the optimal model capable of accurately predicting C─H bond energies was selected. This model also further revealed the dominant electronic structural features of the predictive model through SHapley additive interpretability (SHAP) analysis. This work establishes an interpretable, data-driven methodology for designing high-efficiency multicomponent catalysts.