fig3

Figure 3. Performances of the regression model and validation procedure on the AFLOW dataset. A, Model performances of eight regression models based on the test set. B, Comparison between the predicted band gap by LightGBM and the calculated PBE band gap from the test dataset. The red circles display the two worst-fitting samples. C, Distribution of band gap in the regression dataset, which shows the right-skewed characteristic. D, SHAP analysis for the feature importance of the LightGBM regression model. E, Confusion matrix of stacking classification model based on the AFLOW dataset. F, Comparison between the predicted band gap and the calculated PBE band gap from the AFLOW database. G, Comparison among the bandgaps of several representative oxides obtained by experiments, PBE-functional calculation, and the LightGBM prediction. AFLOW: Automatic Flow of Materials Discovery Library; PBE: Perdew-Burke-Ernzerhof; SHAP: Shapley Additive exPlanations.