Interpretable machine learning‐assisted high‐throughput screening of highly active nitrogen fixation dual‐atom catalysts
Qiang Liu, Yi Wei, Lintao Xu, Yongan Yang, Jingnan Wang, Xi Wang
Abstract
Abstract Machine learning (ML) has addressed the traditional challenges of large data processing in density functional theory (DFT) calculations. However, understanding the relationship between fundamental descriptors and catalytic performance and identifying key drivers of catalytic activity remain challenging. Here, we present a cost‐effective, high‐throughput, and interpretable ML method to accurately identify nitrogen reduction reaction (NRR) performance determinants. Initially, 378 M 1 M 2 @TiO 2 catalysts are screened, yielding 33 promising candidates through high‐throughput techniques. Subsequently, ML models (primarily XGBoost) predict free energy changes of key NRR intermediates. Shapley Additive Explanations (SHAP) analysis identifies two critical features: the M 1 NN bond angle (M 1 NN) and the M 2 N bond length. Four catalysts exhibiting energy changes below 0.3 eV in the potential‐determining step are identified as promising candidates. Combined SHAP analysis and electronic structure calculations confirm the inherent activity of NRR catalysts, highlighting the importance of fundamental properties in modulating active sites for superior NRR performance.