Machine learning predictions of band gap and band edge for (GaN)1-x(ZnO)x solid solution using crystal structure information

The (GaN)(1-x)(ZnO)(x) solid solution is an ideal material for the next generation photocatalyst due to good chemical stability and excellent optical property. Although full range content regulation of ZnO has been achieved, the isomeric phenomena of solid solutions make it difficult to establish a structure-property relationship. Here, we constructed a series of random (GaN)(1-x)(ZnO)(x) structures and calculated the band properties using DFT. Seven supervised machine learning models were trained to understand band properties base on microstructure. The results show that the Random Forest Regressor model is optimal for predicting band gap and band edge position with proposed microstructure descriptors. Feature importance and SHAP analyses indicate four local microstructures are main structural factors influencing band structure. This work is helpful for understanding the relationship between microstructure and band property, and designing excellent photocatalytic (GaN)(1-x)(ZnO)(x) solid solutions.

Automated summary

In this study, a series of random (GaN)(1-x)(ZnO)(x) structures were constructed and seven machine learning models were trained to predict their band properties. The Random Forest Regressor model was found to be optimal for predicting band gap and band edge position. Feature importance and SHAP analyses revealed that four local microstructures are the main factors influencing the band structure. This work contributes to the understanding of the relationship between microstructure and band properties, and aids in the design of excellent photocatalytic (GaN)(1-x)(ZnO)(x) solid solutions.