Machine-Learning-Guided Genetic Inverse Design of Single-Atom Electrocatalysts for CO2 Reduction
The rational design of materials with tunable compositions remains challenging due to the vastness of compositional space. Herein, we propose a genome-inspired materials intelligence framework (GIMI) for inverse design in high-dimensional compositional spaces. By integrating a multi-estimator disagreement-based data filtering strategy with a genetic algorithm, this framework enables on-the-fly improvement of the machine-learning predictive accuracy and efficient exploration of diverse compositions, thereby significantly enhancing search efficiency while reducing computational cost. Applied to graphene-based single-atom catalysts (SACs) with variable ligands for CO2 electroreduction to CO, GIMI efficiently screens 34,992 possible metal-ligand combinations and identifies promising SACs (e.g., Zn-O1N3 and Zn-O2N2) by evaluating only ∼1250 structures per round, demonstrating its high search efficiency. Further interpretability analysis reveals that the cohesive energy and electronegativity of the metal center primarily govern CO2RR activity, while ligands play a secondary role by modulating the local coordination geometry. This work establishes a scalable and generalizable platform for inverse materials design, enabling targeted exploration of complex compositional space and accelerating the discovery of high-performance catalysts.