Machine-Learning Discovery of Metal–Organic Framework Crystals for Second-Harmonic Generation

Abstract Nonlinear optical (NLO) crystals are critical for frequency conversion, light generation, and nonlinear photonics. Among inorganic and organic NLO crystals, hybrid metal–organic frameworks (MOFs) are emerging as promising crystalline alternatives owing to their high polarizability, endurance, and structural tunability. However, the latter leads to enormous compositional and structural diversity among MOFs (>130,000 reported structures), making the rational discovery of high-performance NLO MOFs challenging using experimental and computational approaches. Here, we report a unified dataset of 2327 crystals: 937 non-centrosymmetric (including 67 experimentally validated NLO MOFs and 870 inorganic NLO crystals) and 1390 centrosymmetric crystals as a reference. Based on this dataset, we develop a machine-learning (ML) model for screening MOFs and identifying promising non-centrosymmetric NLO candidates for second-harmonic generation (SHG). The model predicts SHG efficiencies with an overall accuracy of 2.05 pm V–1 (mean absolute value), and 0.39 pm V–1 for the experimental MOF dataset, and ranks non-centrosymmetric crystals with a Spearman correlation of 0.66. Screening is 105 times faster than DFT, positioning the model as a ranking filter for prioritizing candidates. This work establishes a data-driven ML model for accelerating the discovery of next-generation NLO MOF crystals and expanding the design space for advanced nonlinear photonics.