Accelerated Photocatalytic C–C Coupling via Interpretable Deep Learning: Single-Crystal Perovskite Catalyst Design using First-Principles Calculations

Photocatalytic C–C coupling reactions have garnered significant attention for their potential to drive sustainable chemical transformations. The design of efficient photocatalysts is critical in optimizing these reactions. In this study, we use a computational materials science approach, leveraging first-principles calculations to evaluate the bandgap values of 158 single-crystal perovskite materials. We employ a deep learning model, incorporating a multi-head-attention mechanism within a ResNet architecture, to predict the bandgap based on features such as τ, Group-A, Group-B, Pettifor number, χM-B, χP-B, Ea-A, cB, KB, and Ra-B. This model's performance is compared to traditional machine learning techniques, including K-means, MLP, Random Forest, PCA, and Multivariable Linear Regression. The results demonstrate that the self-attention ResNet model achieves a training R2 of 0.819 and a test R2 of 0.803, indicating strong predictive accuracy. The model’s interpretability is enhanced by visualizing the permutation importance of each feature, shedding light on the contributions of various factors to the prediction. These findings highlight the potential of machine learning, particularly deep learning, in accelerating the design of photocatalysts for C–C coupling reactions.