Exploration of New Materials for Nonlinear Optical Crystals using Materials Informatics

Recently, materials design approaches that introduce materials informatics have become extremely important in materials science. MI methods can accelerate the design of new materials by using machine learning to predict physical properties with accuracy comparable to density-functional-based first-principles calculations and with computational speeds an order of magnitude faster.<br/>Among them, the Atomistic Line Graph Neural Network (ALIGNN) [1], which considers distance and angle information, has been shown to significantly improve feature representation between atoms and prediction accuracy. However, there is a tradeoff between prediction accuracy using deep learning such as ALIGNN and the number of training data. For many practical materials, it is difficult to obtain sufficient prediction accuracy because of the small amount of property data that is available, due to computational costs.<br/>In this study, we introduce transfer learning into ALIGNN to construct a sufficiently accurate property prediction model on a small data set by pre-training on a large amount of property data already obtained from databases and other sources.<br/>Nonlinear optical materials are promising materials that have an important role in expanding the wavelength range of lasers.<br/>Among them, there is an urgent need to develop ideal materials that simultaneously satisfy a wide band gap, a large second harmonic generation (SHG) susceptibility, and an appropriate birefringence. Nonlinear optical susceptibility and birefringence can be obtained by electric field calculations using density functional perturbation theory, but they remain computationally expensive.<br/>Therefore, in this study, we predict nonlinear optical susceptibility and birefringence from a small data set by using ALIGNN and transfer learning.<br/>ALIGNN and transfer learning were implemented in PYTORCH [2], a Python library. First, nonlinear optical properties were calculated with the CASTEP code [3] for about 500 structures obtained from the Inorganic Crystal Structure Database (ICSD) [4] including B and Ge. Then, thousands to tens of thousands of structural data and properties (band gap, refractive index, DOS) data that are close to the composition of the 500 structures were obtained from the Materials Project Database (MPD) [5] as pre-training data and prediction models were constructed for each of them. We predicted the nonlinear optical susceptibility and birefringence by training only the final layer of the optimized ALIGNN model. Finally we screened from various compositions and structures by using the predictive models of nonlinear optical properties constructed.<br/><br/>[1] Choudhary, K., DeCost, B. Atomistic Line Graph Neural Network for improved materials property predictions. Npj Comput Mater 7,185(2021).<br/><br/>[2] Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., … Chintala, S. (2019). PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems 32 (pp. 8024–8035).<br/><br/>[3] "First principles methods using CASTEP", Zeitschrift fuer Kristallographie 220(5-6) pp. 567-570 (2005) S. J. Clark, M. D. Segall, C. J. Pickard, P. J. Hasnip, M. J. Probert, K. Refson, M. C. Payne<br/><br/>[4] NIST Inorganic Crystal Structure Database, NIST Standard Reference Database Number 3, National Institute of Standards and Technology, Gaithersburg MD, 20899<br/><br/>[5] Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, Kristin A. Persson; Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater 1 July 2013; 1 (1): 011002.