Accurate prediction of semiconductor bandgaps based on machine learning and prediction of bandgaps for two-dimensional heterojunctions

The bandgap value of materials has a profound impact on their properties and applications. Presently, with the development of high-throughput calculations, the bandgap of most materials is simulated and calculated using the density functional theory (DFT). Nevertheless, the bandgap of materials calculated in this way is often accompanied by large errors and long time consuming. Besides, the bandgap results obtained in different experimental environments are different. Therefore, finding a method to calculate the material bandgaps quickly and accurately is imminent. In this work, the Machine Learning (ML) method is used to predict the bandgap of semiconductor materials. Four different machine learning models are trained and tested through the feature processing, which can accurately predict the bandgap of the material by the hybrid density functional (HSE06) method, of which the average mean absolute error (MAE) is 0.142 eV and the coefficient of determination (R2) is 0.977. Moreover, in order to better predict the bandgap of local small-sample semiconductor materials, the federal learning framework is employed to forecast small-sample datasets under different experimental condi-tions. Then, the four ML models are used to the prediction of materials and compared the results with the local calculation results. The results indicate that the bandgap error of compound semiconductor materials is 2-10%, and the bandgap error of 2D heterojunctions semiconductor materials is 5-30%. In addition, the ML models are also utilized to the Materials Project database, in which the bandgap of about 53170 semiconductor materials is successfully predicted. In conclusion, the work not only provides a method to accurately predict the bandgap of compound semiconductor materials, but also supplies an effective idea for the prediction of the bandgap of semiconductor materials in local small data sets, which accelerating the development of the application of semiconductor materials.

Automated summary

The bandgap of semiconductor materials can be accurately predicted using machine learning methods, with an average absolute error of 0.142 eV and a coefficient of determination of 0.977 achieved through feature processing using density functional theory. The federal learning framework is employed to forecast bandgap under different experimental conditions for small-sample datasets, with bandgap errors of compound semiconductor materials at 2-10% and 2D heterojunction semiconductor materials at 5-30%.