Machine Learning with Multilevel Descriptors for Screening of Inorganic Nonlinear Optical Crystals

Nonlinear optical (NLO) crystals are the key materials in modern laser technology and science because of their intrinsic capability to convert the wavelength of the light source. The search for new NLO materials is still very active in both scientific and industrial communities. Machine learning (ML) becomes a powerful tool to explore new candidates of NLO materials and to reveal the underlying relationship between structures and properties. In this work, we have proposed multilevel features that are relevant to the atomic properties, the characters of fundamental structural groups, and the crystal structures to describe inorganic NLO crystals for machine learning. The first-level and second-level descriptors can be obtained based on chemical compositions of crystals without prior knowledge about crystal structures. Several ML classifiers have been optimized using a database that consists of hundreds of NLO crystals to identify the samples with desired birefringence (Δn) and second-order nonlinear coefficients (dij). In particular, almost all of the ML models that only involve the first-level and second-level features, called as the crystal-structure-free model, exhibit good classification performance. It is still far from perfect but suitable to act as a filter in the first step of high-throughput materials discovery. Using the optimized ML models, feature importance analyses and virtual screening processes have been performed to understand the relationship between the features and targeted properties and to extract the statistical pictures on elements and fundamental structural groups. Several unexplored crystals are also picked out as ML-proposed candidates, and three of them are suggested as new potential NLO materials based on further first-principle calculations. The present ML models are expected to accelerate the inverse design for new NLO crystals with desired properties.