Fast-Track to Catalyst Stability: Machine Learning Optimized Predictions for M1/M2-N₆-Gra Catalysts

Graphene-based dual-atom catalysts M1/M2-N₆-Gra have shown significant potential in various reactions, although their stabilities are debated. Therefore, developing an efficient and accurate approach to screen thermodynamically stable M1/M2-N₆-Gra is significant. Herein, we designed a rational machine learning (ML) scheme based on 143 DFT calculated samples to predict the formation energies (<i>E</i>_<i>f</i>) of 1134 possible M1/M2-N₆-Gra. A well performing multilayer perceptron model with test set <i>R</i>² <i>=</i> 0.98 was obtained after feature engineering, model training, data supplementation, and transfer learning. This model successfully screened 604 thermodynamic stable M1/M2-N₆-Gra with <i>E</i>_<i>f</i> < 0 eV. Feature importance, predictions distribution, and energy decomposition revealed that the coordination number significantly influences <i>E</i>_<i>f</i>, with cohesive energy dominating low-coordination catalysts and binding energy between metal and substrate being more critical in higher-coordination catalysts. This work highlights the potential of ML and developed effective approaches to screen thermodynamically stable catalysts and reveals the laws of stability for various materials.