Infrared nonlinear optical (NLO) materials are essential for laser and photonic technologies, limited by fragmented material systems, lengthy development cycles, and trial-and-error synthesis. To overcome these barriers, we developed an integrated computational-experimental framework integrating first-principles high-throughput calculations, machine learning, and targeted synthesis. We establish a multidimensional properties dataset of 1807 non-centrosymmetric compounds and define a comprehensive figure of merit (CFOM) Q based on the statistical average of this dataset to quantify performance trade-offs. Multidimensional statistical analysis uncovers composition-structure-performance relationships, and reveals superior structure and chemical compositions governing enhanced NLO performance. A Q-based crystal graph neural network classifier is developed, achieving strong predictive accuracy (AUC = 0.95). We identify 12 unreported candidates (Q > 2) from 5105 compounds combining high-throughput calculation and machine learning. Experiments confirm that defect-chalcopyrite HgAl2Q4 (Q = S, Se, Te) shows wide band gaps (1. 55-2.82 eV), suitable birefringence (0.06-0.08), and strong NLO responses (2.2-5 × AGS). This work provides an effective pathway for accelerating the discovery of high-performance optoelectronic materials.
Xiao et al. (Fri,) studied this question.