A tool to predict the evolution of phase and Young’s modulus in high entropy alloys using artificial neural network

Abstract

We report an artificial neural network (ANN) based tool to predict the evolution of phase(s) in high entropy alloys (HEAs) and their Young’s modulus. Two independent networks as single channel output and four channel binary output methods have been adopted for testing as well as for predicting the evolution of either single-phase BCC, FCC solid solution (SS) phase, or a mixture of SS phase(s) and intermetallic compound (IM). The data of as-solidified HEAs were collected from the literature, and various thermo-physical as well as electronic parameters, such as mixing entropy (ΔSmix), mixing enthalpy (ΔHmix), atomic size difference (δr), Allen electronegativity ΔχA, Pauling electronegativity (ΔχP), and valence electron concentration (VEC) were calculated for each alloy composition, which was used as input features. The Bayesian regularization backpropagation has been adopted as the learning algorithm. Furthermore, the model has been trained and validated by developing eutectic CoCrFeNiNbx (x = 0.45, 0.5, 0.55 and 0.65) and CoCrFeNiTay (y = 0.2, 0.4, 0.455 and 0.5) HEAs experimentally. The single channel output approach and four channel binary output approach generate the matching accuracy of 85.95% and 92.97% for the main dataset, where the same is 70.83% and 91.67% for the deployment dataset, respectively for phase prediction. The predicted values of Young’s modulus have achieved the relative accuracy of 94.96% and 96.44% for the main and deployment dataset, respectively. TheANN suggests that VECbecomes the dominant factor for phase stability when δr < 3%, whereas the higher value of δr promotes the evolution of two phases in HEAs.