A Yield Strength Prediction Framework for Refractory High-Entropy Alloys Based on Machine Learning

Abstract

Machine learning has been widely applied to materials research with the development of artificial intelligence. Here, a new framework mainly based on the LightGBM algorithm was proposed, which predicted the yield strength of refractory high-entropy alloys (RHEAs) in various temperatures. The features of T, D.B, μ, Smix, Gmix and r were recognized as the optimal feature set by several feature screening methods. The framework displayed good prediction results with a coefficient of determination (R2) of 0.9605 and a root mean square error (RMSE) of 111.99 MPa in the test set. A series of RHEA samples validated the generalization of this framework. SHAP with pearson correlation constant (PCC) and maximal information coefficient (MIC) interpreted the framework and analyzed the intrinsic mechanism of features on yield strength, discovering a novel μ-D.B-Gmix design strategy for obtaining RHEAs with enhanced yield strength. Both TiTaNbHfNi0.25 and TiTaNbHfNi0.5 alloys were fabricated as the experimental verification for this framework which showed 1230 and 1311 MPa yield strength with the predicted errors of 6.3% and 3.7%. The validations above demonstrated the excellent performance of the present framework and the effectiveness of such a strategy.

Keywords: refractory high-entropy alloys, Machine learning, Yield strength, Computational model, Alloy design