Wang, Lin, Li, Peiyou, Zhang, Wei, Wan, Fangyi, Wu, Junxia, Yong, Longquan, and Liu, Xiaodi
To avoid the lack of unified physical significance of the random combination of characteristic parameters, four characteristic parameters with potential energy distribution were selected to predict the phase selection of three type of amorphous alloys (AM), solid solution alloys (SS) and high entropy alloys containing intermetallic compounds (IM) by artificial neural network (ANN) in machine learning. To improve the prediction accuracy, the combination of three different parameters can be used to predict the phase of AM and IM alloys, and the combination of four different parameters can be used to predict the phase of SS alloys. The mean square error (MSE) describes the error between the real value and the predicted value, which directly affects the prediction accuracy. For the AM and IM alloys, the partial three parameter combinations with the lowest MSE values have highest prediction accuracy, for SS alloys, the four parameter combination with the lowest MSE value has highest prediction accuracy. The correlation coefficient (R) is used to evaluate the fitting effect of the model. Based on the correspondence between the R values and the prediction accuracy, it can be concluded that the current ANN model is accurate in predicting the phase selection of three type of alloys, and is the good learning model. The sensitivity matrix (S) calculated according to the weight (w) is an important parameter that affects the prediction accuracy. The S values of the atomic size difference (δ) have a greater impact on the phases of AM, SS and IM alloys; however, the corresponding S values of mixing enthalpy (Δ H m) in the AM and SS alloys have the weak influence. [Display omitted] • Four parameters were selected to predict phases of three types alloys. • Three parameter combination were used to predict phases of AM and IM alloys. • For AM and IM alloys, partial three parameter combinations have highest prediction accuracy. • Current ANN model is accurate in predicting phases of three type of alloys. To avoid the lack of unified physical significance of the random combination of characteristic parameters, and to solve the problem that there are many factors influencing the phase selection by multiple parameters, four characteristic parameters with potential energy distribution were selected to predict the phase selection of three type of amorphous alloys (AM), solid solution alloys (SS) and high entropy alloys containing intermetallic compounds (IM) by artificial neural network (ANN) in machine learning. To improve the prediction accuracy, the combination of three different parameters can be used to predict the phase of AM and IM alloys, and the combination of four different parameters can be used to predict the phase of SS alloys. For the AM and IM alloys, the partial three parameter combinations with the lowest mean square error (MSE) values have highest prediction accuracy, for SS alloys, the four parameter combination with the lowest MSE value has highest prediction accuracy. Based on the correspondence between the correlation coefficient (R) values and the prediction accuracy, it can be concluded that the current ANN model is accurate in predicting the phase selection of three type of alloys, and is the good learning model. The sensitivity matrix (S) values indicate that the atomic size difference (δ) have a greater impact on the phases of AM, SS and IM alloys; however, the corresponding S values of mixing enthalpy (Δ H m) in the AM and SS alloys have the weak influence. The current learning model and the combination of three or four characteristic parameters can predict the AM and SS phase varified by X-ray diffraction of new Ti-Cu-Ni-Zr (AM) and Fe-Co-Ni-Cu-Ti (SS) alloys, thus accelerating the composition design and phase composition selection of new alloys. [ABSTRACT FROM AUTHOR]