Symmetry‐Aware Equivariant Network for Discovering SHG‐Active Materials

Abstract The discovery of novel materials with exceptional nonlinear optical (NLO) properties, especially high second‐harmonic generation (SHG) coefficients, is vital for advancing optical communication technologies and solid‐state laser systems. Computational and machine learning methods open up enormous opportunities for accelerating the identification and characterization of new SHG‐active materials. In this study, OptiXNet, a physics‐informed deep learning model leveraging symmetry‐aware equivariant graph neural networks is introduced, specifically designed to achieve unprecedented accuracy in predicting SHG tensors. Trained and validated on the NOEMD database, OptiXNet significantly outperforms existing state‐of‐the‐art methods in both prediction accuracy and generalization to unexplored chemical compositions. To ensure robust predictions when exploring unknown chemical spaces, an automated clustering algorithm is integrated that quantifies model confidence and flags out‐of‐distribution data points. Using this clustering‐guided approach, MgP 2 O 5 F 2 —a previously unexplored material is successfully identified with promising SHG properties—within the model's high‐confidence predictive domain. Subsequent confirmatory density functional theory (DFT) calculations validated the accuracy and effectiveness of the approach. Combining symmetry‐aware deep learning, rigorous uncertainty quantification, and reliable computational validation, the framework provides a powerful and efficient strategy for the accelerated discovery of next‐generation NLO materials.