Data-Driven Machine Learning with Lattice Distortion and Thermodynamic Parameters Guided Strength Optimization of Refractory High-Entropy Alloys

Abstract

The development of refractory high-entropy alloys (RHEAs) through conventional trial-and-error approaches I s highly inefficient given the vast compositional space. To address this difficulty, a data-driven machine learning (ML) model was established to predict the compressive yield strength (σ0.2) of RHEAs, which provides a design strategy with two pivotal descriptors concerning lattice distortion and thermodynamic parameters. The LightGBM algorithm was utilized to train the model based on a RHEA yield strength dataset. The optimal descriptor set consists of T, Gmix, D.B, VEC, μ, δr, and the resultant model obtained a determination coefficient (R2) of 0.88 and a mean absolute error (MAE) of 0.162 GPa in 10-fold cross validation, exhibiting remarkable accuracy and generalization. Interpretation methods clarified the influence of each descriptor. As a result, a linear equation was revealed as a decision boundary for the rapid design of RHEAs with high σ0.2. The effects of incorporating Ni and Al into the TiMoNbHf, TiTaNbHf, and TiZrNbHf matrix RHEAs on σ0.2 were investigated. Experiments verified the application of such a model and the strategy with a mean error of 7.2%, and the designed TiMoNbHf(AlNi)0.2 displayed σ0.2 up to 1.61 GPa with fracture strain of 20%.

Keywords: Refractory high-entropy alloys, Machine Learning, Yield strength, materials design