Accelerating the discovery of athermal deep-ultraviolet nonlinear optical crystals via physics-informed herarchical graph learning

Thermal-induced frequency instability significantly hinders the performance of deep-ultraviolet (DUV) coherent sources. We propose a passive intrinsic control strategy using a physics-informed hierarchical graph learning framework to discover athermal nonlinear optical (NLO) crystals. By integrating anisotropic attention and space-group embeddings with physics-informed feature injection, the model achieves high predictive fidelity, with R 2 exceeding 0.95 for electronic bandgap and formation enthalpy, and a Spearman’s rank correlation coefficient ( ρ ) above 0.90 for the second-harmonic generation (SHG) response. Virtual screening on the Novel Opto-Electronics Materials Database (NOEMD) demonstrates a 7.91-fold enrichment factor within the top-30 candidates, yielding an effective discovery rate of 73.3%. Interpretability analysis indicates that the framework can capture structural features and physical regularities consistent with established chemical heuristics. This work provides an efficient paradigm for the data-driven design of multi-functional photonic materials for stable DUV frequency comb generation.