Features engineered hardness and yield strength prediction of septenary refractory high-entropy alloys

Refractory high-entropy alloys (HEAs) with more principal elements could have better mechanical properties, however, they also face greater challenges for designing because of larger exploration space of mechanical properties and complex physical interactions among elements. In this work, a machine learning (ML) model for the prediction of hardness and yield strength of septenary RHEAs was built. By comparing four ensemble models for 5-fold cross-validations, the CatBoost model was finally selected due to its better prediction performance. To overcome the shortcomings with limited datasets for RHEAs with increased number of principal elements, feature engineering was applied to expand the existing ordinary features. Multiple feature selection methods were combined in order to retain features that had a critical effect on hardness and yield strength. Pearson correlation coefficient was used to assess the degree of linear correlation among features. Thereafter, Shapley Additive Explanations (SHAP) was used to analyze the impact of each feature on prediction. Based on feature engineering, four RHEAs were recommended by the CatBoost model, and their mechanical properties were characterized. The results showed high accuracy for the hardness and yield strength prediction of septenary W-Nb-V-Zr-Cr-Mo-Ti RHEAs, where the WNb 2 V 2 Zr 2 Cr 2 MoTi alloy even demonstrated mean relative errors (MREs) of 0.18 % and 1.13 % for hardness and yield strength respectively. The present findings confirmed the effectiveness of feature engineering on the prediction of RHEAs with increased number of principal elements using ML models.