Machine Learning Prediction on Birefringence of Nonlinear Optical Crystals and Polymorphs with Different Birefringence Activities

Nonlinear optical (NLO) crystal materials have been widely used in the scientific and industrial fields. Birefringence is an important property of the NLO crystals. Tuning appropriate birefringence through element substitution or polymorphic transformation may promote phase-matching performance facing various demands of laser wavelength. A growing number of studies based on machine learning (ML), such as the multilevel descriptors developed in our group (Zhang et al. <i>J. Phys. Chem. C</i> 2021, 125, 25175-25188), can successfully predict birefringence of NLO materials. However, how to identify polymorphs with different birefringence activities is still a nascent research topic. In this work, we proposed hp-wACSFs, a new descriptor based on the widely used atom-centered symmetric function, to predict the birefringence of inorganic crystals. A series of ML classifiers were built using hp-wACSFs. Two learning tasks, which aim at birefringence-active NLO crystals or polymorphs with different birefringence activities, were implemented. The performance on the former task was as good as our previously reported work, while the best accuracy on the latter task, which cannot be processed in the absence of three-dimensional descriptors, achieved 0.8 in this work. We finally implemented virtual screening using constructed ML models to search polymorphs with different birefringence activities.