A hybrid prediction frame for HEAs based on empirical knowledge and machine learning

Phase formation plays key role in the properties of high-entropy alloys (HEAs). If the phases of HEAs can be accurately predicted, the number of experiments can be greatly reduced, and the process of material design can be greatly accelerated. Machine-learning methods have been successfully and widely applied to predict the phases of HEAs. However, the accuracy of a single machine-learning (ML) algorithm is not ideal and different ML algorithms may predict different results. These issues hinder the application of ML in material design. In this paper, a hybrid frame for HEAs phase prediction, which combines machine-learning and empirical knowledge, is proposed. First, for the purpose of solving the problem that a sample may be predicted as inconsistent prediction phases by different algorithms, the Dempster-Shafer (DS) evidence theory is adopted to fuse the inconsistent of the predicted phases among different algorithms, and provide a fusion prediction phase with the highest credibility. Second, a conflict-resolution model with high accuracy based on the improved DS evidence theory is proposed. Last, the empirical knowledge criterion is combined with the conflict-resolution model to improve the efficiency and accuracy of the hybrid prediction frame. The 426 different HEAs samples consisting of 180 quinaries, 189 senaries, and 57 septenaries were collected to validate against the effectiveness of the proposed methods. The experimental results demonstrate the hybrid prediction frame achieves higher accuracy and better performance than single ML algorithm. Keywords: Hybrid model; High-entropy alloys; Phase prediction; DS evidence theory