Prediction of Phase Structures of High-entropy Alloy Based on Feature Engineering and Machine Learning

788 high-entropy alloy data sets were obtained through removing unreasonable and redundant data by data filtering, 13 characteristic parameters for predicting phase structures were summarized, and the redundancy was removed by Pearson correlation coefficient method. The order and quantity of feature input models were determined by feature removal method. Four standardized preprocessing methods of feature data sets were compared, including regularization, robust standardization, range normalization and standard deviation standardization. three kinds of algorithm of random forest, support vector machine and nearest neighbor regression were compared. the best preprocessing method and algorithm were determined by cross-validation, and the machine learning model was established. the model was used to predict the high-entropy alloy phase structure. The results show that considering the accuracy and efficiency of the model comprehensively, the standard deviation of melting point, standard deviation of valence electron concentration, atomic radius difference, mixing entropy, average melting point of alloy component elements, valence electron concentration, mixing enthalpy, electronegativity difference and average atomic radius were selected as the feature set, and was input into the model in this order. The best feature data set preprocessing method was standard deviation standardization, and the best algorithm was random forest algorithm. The accuracy of the constructed model for predicting intermetallic compound, amorphous phase, face-centered cubic phase and body-centered cubic phase structures was 95%, 89%, 88% and 90%, respectively. The accuracy of predicting face-centered cubic + body-centered cubic complex phase structure and body-centered cubic + intermetallic compound complex phase structure was about 76%; the accuracy was generally high, verifying the generalization ability of the model.