The phase design of high entropy alloys (HEAs) is an important issue since the phase structure affects the comprehensive properties of HEAs. Accurate prediction of phase classification can accelerate material design. In this paper, a new phase prediction framework is proposed using machine learning (ML) and an improved information fusion approach based on the Dempster-Shafer (DS) evidence theory. Considering that the classification results of different ML algorithms may conflict, and the traditional DS evidence theory cannot solve the problem of high conflict, we propose an improved information fusion approach based on the DS evidence theory. The basic probability assignment function is constructed using the ML algorithms. 761 HEAs samples are collected consisting of amorphous phase (AM), solid solution (SS), intermetallic compound (IM), and a mixture of SS and IM (SS + IM). For the small dataset of HEAs, we use a conditional generative adversarial network (CGAN) for data augmentation. Based on the enhanced dataset, the ML model is optimized by sparrow search algorithm (SSA), which can accelerate searching speed of model hyperparameters and improve the performance of the model. The results show that the proposed information fusion method performs better than several other existing techniques on the test set, and the prediction accuracy reaches 94.78 %. Meanwhile, the prediction accuracy of the proposed method is higher than that of the existing technology (93.17 %). It is proved that the proposed method can solve the high conflict problem effectively. Moreover, we present the interpretability analysis of the features by the Shapley additive explanations (SHAP) and the sensitivity matrix. A smaller atomic size difference δ (<6.6 %) is conducive to the formation of SS phase, while a larger δ (>6.6 %) is conducive to the formation of AM phase. A smaller enthalpy of mixing ΔHmix tends to form AM phase. In binary and ternary alloy systems, IM phase can be extracted by the mixing enthalpy ΔSmix < 10. In addition, we find that mean bulk modulus (K) and standard deviation of melting temperature (σT) are critical features to distinguish between SS and SS + IM.
Read full abstract