Accelerating the Prediction of g‐C 3 N 4 ‐Supported Dual‐Atom Catalysts for Photocatalytic CO 2 Reduction to CO and HCOOH: A Machine Learning and DFT Combined Approach

Abstract Dual‐atom catalysts (DACs) show promise for photocatalytic carbon dioxide (CO 2 ) reduction due to their high atom utilization and synergistic effects. However, finding efficient combinations is challenging because of the large number of possibilities. In this study, a high‐throughput screening of TM1TM2@g‐C 3 N 4 catalysts is conducted using density functional theory (DFT) and machine learning (ML) to identify promising candidates for CO 2 photoreduction. The results reveal that the ML algorithm can successfully achieve the relationship between the descriptors of the DACs and the limiting potentials ( U L ) of CO and HCOOH products. Among four ML models, the feedforward neural network (FNN) achieves the highest accuracy. The ML model successfully predicts the limiting potentials ( U L ) and screens RuHf@g‐C 3 N 4 ( U L = −0.72 V) and VV@g‐C 3 N 4 ( U L = −0.79 V) as promising CO producing catalysts, and MoMo@g‐C 3 N 4 ( U L = −0.28 V) and CrMo@g‐C 3 N 4 ( U L = −0.32 V) as efficient HCOOH producing catalysts. DFT validation shows low average errors (−0.05 eV for CO, 0.02 eV for HCOOH). Compared to pure DFT, the FNN model reduces screening time by 56% (CO) and 42% (HCOOH). The ML framework not only successfully screens highly promising catalysts but also provides a solid theoretical basis for the subsequent experimental synthesis for energy conversion.