First-principles calculations steered multi-task transformer model to screen dual-atom catalysts for C-H activation

Activating C-H bonds in light alkanes (methane, ethane, and propane) remains a major challenge in heterogeneous catalysis due to their high stability, while dual-atom catalysts (DACs) offer a promising balance of activity and durability. Herein, we developed an optimized multi-task transformer model to accelerate active DACs design by integrating first-principles calculations with machine learning. Over 200 DACs composed of fourth/fifth period transition metals anchored on N-doped graphene were studied. The transformer, modified with additional decoder layers and attention head, outperformed conventional methods (K-nearest neighbors, resource description framework, and gradient boosting regression) for small datasets, achieving R² > 0.85 in predicting both adsorption energies and C-H activation barriers. Gradient boosting regression tree analysis identified the metal-N coordination distance as the key performance-determining factor. This work provides an accurate predictive model for efficient DACs screening in light alkane conversion and advances DACs design principles.