Predictive nonlinear optical crystal formation energy regression model based on convolutional neural networks

Abstract

In modern laser science, nonlinear optical crystals play a crucial role as indispensable functional materials. Overcoming the inherent defects of traditional materials such as two-photon absorption and low laser damage thresholds, the search for high-performance nonlinear optical crystals remains an urgent and unresolved challenge. In this study, we extensively utilized deep learning techniques to construct a lightweight convolutional neural network. This network is built on the basis of information about the nuclear charge of elements and the number of atoms in compounds, creating carefully designed feature maps for accurately predicting the formation energy of nonlinear optical crystals. Our model demonstrates outstanding performance, with predictive metrics showing R² = 0.985, RMSE = 0.128, and MAE = 0.083. This remarkable predictive capability enables our network to efficiently predict the formation energy of crystals and reliably assess their stability. Our approach is poised to play a crucial role in high-throughput screening, significantly reducing computational burdens and accelerating the discovery process of novel crystals. This significant achievement provides an efficient and reliable tool for crystal material research, opening up new possibilities for future materials science investigations.