Machine Learning Regression Model for Predicting the Formation Energy of Nonlinear Optical Crystals

Abstract

AbstractNonlinear optical crystals have been a key material in lasers due to their excellent frequency conversion properties. The search for new nonlinear crystals has been extremely active due to the inherent drawbacks of existing nonlinear crystals, such as the low laser damage threshold of infrared nonlinear crystals and the phenomenon of two‐photon absorption. The stability of a material can be characterized by calculating the formation energy, and machine learning (ML) is currently the most mainstream tool to explore materials. In this paper, the digital features of nonlinear crystals are established by using only the component information. Based on the low‐variance and cross‐validated recursive feature elimination (RFECV), 11 features with strong correlation are obtained for model training. Among them, the two features related to electronegativity account for 71.44% of all features, and become the most critical factor for the formation energy predicted by the model. Three ML regression models have been proposed to predict the formation energy of nonlinear crystals. Among them, the gradient boosting regression model shows excellent prediction performance with R2 = 0.939, RMSE = 0.205, and MAE = 0.132 eV per atom. The model provides a useful tool for fast and low‐cost prediction of nonlinear crystal formation energy.