Machine Learning Study On Rare Earth-based Amorphous Alloys

The magnetic refrigeration technology based on magnetocaloric effect(MCE)has been considered as a more advantageous technique over present well-used conventional gas compression.However,the composition of a vast number of magnetic materials with large MCE are still remained unknown.On the other hand,data-mining techniques using machine learning are found really efficient in many fields,especially for materials design.So here,due to the distinctive advantaged of amorphous alloy over crystalline materials and the intrinsic large MCE of heavy rare earth-based alloys,the system of rare earth-based amorphous alloy was selected to study by this technique.After trying several machine learning algorithms,the gradient boosting regression tree(GBRT)ensemble algorithm with the best performance was selected.And by putting the external magnetic field into input features together with the composition,two models were successfully built to predict Curie temperature and magnetic entropy change with high accuracy.The performance metric coefficient scores of determination(R~2)of the two models are 0.96and 0.92.Considering the error that exists in the magnetic entropy change itself,a new standard was presented to observe the prediction results for magnetic entropy change and found that almost all of the deviation from the actual values was allowable.It was proved that both of the two models possess excellent generalization ability and can be actually applied to guide the experiments to synthesize rare earth-based amorphous alloys with large MCE.And due to the particularity of the system selected here,before actual application,the glass forming ability(GFA)of materials should be studied.Meanwhile,people can't fully understand and predict GFA and this problem still greatly limits the application of metallic glass with excellent properties in industrial field.So in this work,the proposed model uses the random forest classification method to solve the GFA prediction for binary metallic alloys.Compared with the previous SVM algorithm models,this new model not only shows the degree of each input feature influence on GFA and the correlation with GFA,but also further guides the author to find a new combination of input features with higher accuracy by analyzing the importance of each input feature.Considering the limitation of the evaluation indicator in the SVM model,a normalized one of binary alloy for machine learning model performance is put forward.Finally,the mode shows the composition of binary alloys that can form metallic alloy by the melt-spun technique with its prediction.The results can guide the experiments to synthesize metallic alloy with high GFA,and on the other hand can find rare earth-based amorphous alloy system from binary alloy to actually apply our MCE prediction model.