Research on Performance Prediction Method of Refractory High-Entropy Alloy Based on Ensemble Learning

Due to the huge component space of refractory high-entropy alloy, the traditional, experimental “trial and error method” can not meet the design requirements. In order to improve the “trial and error method”, guidance is provided for the prediction and design of refractory high-entropy alloys. Based on the literature data, a comprehensive dataset was constructed, including the composition, phase composition, and strength data of various high-entropy alloys. On this basis, nine regression models were established for strength prediction. By comparison, the XGBoost (XGB) model achieves better prediction performance in the test set; the root mean square error (RMSE) is 195.53 MPa, and the coefficient of determination (R²) is 0.87. By using Shapley additive interpretation (SHAP) to analyze the explainability of the model, it was found that the key characteristics affecting the mechanical properties of the high-entropy alloy were mixed entropy and electronegativity. In order to further evaluate the precision of the model, through the vacuum arc melting preparation, Ti27.5Zr26.5Nb25.5Ta8.5Al12 high-entropy alloys were experimentally verified. The alloy experiments’ yield strength was 1356 MPa, predicting strength was 1304.71 MPa, external validation error was 3.81%, and the average accuracy of the model was 87.38%.