The study of how to “control forming and performance” during the thermal deformation of metal materials has always been a central theme in academic research, particularly in addressing the processing challenges asso...The study of how to “control forming and performance” during the thermal deformation of metal materials has always been a central theme in academic research, particularly in addressing the processing challenges associated with difficult-to-form alloys that possess unique functionalities. However, neither the currently commonly used phenomenological constitutive model, physical constitutive model, Dynamic Material Model (DMM) thermal processing theoretical model, and Ruano-Wadsworth-Sherby (R-W-S) deformation mechanism map model incorporating dislocation density nor the reported machine learning method has established a universal model that can achieve a quantitative description of the process-microstructure-formability of thermal processing. It is only possible first to use modeling research to obtain the law of thermal deformation behavior of alloys and then use the results of microscopic characterization to verify the theory. The research methods lack the characteristics of diagnosis and prediction optimization. This study proposes a machine learning framework for optimizing the random forest (RF) model based on a multivariate decision tree, including microstructure images and hot working process parameter information. It predicts the critical performance parameters, energy dissipation behavior, optimal processing window, and softening mechanism of ternary shape memory alloy Ni_(47)Ti_(33)Hf_(20) in the hot working process. This model has a certain universality. It enables coupled analysis of image information and process parameter data and introduces the calculation and ranking of feature importance, reflecting the applicability of feature values in model construction. Finally, the visualization technique Grad-CAM describes the correlation between the input microscopic image and the output, generating critical hotspots in the heat map. The model of accuracy in predicting the power dissipation rate is confirmed by the grain misorientation angles, thus realizing the establishment of a mechanism-driven model based on the evolution of critical microscopic structures during the thermal deformation of the alloy, which dramatically improves the interpretability of the machine learning model. This machine learning framework provides valuable guidance for quantitatively predicting the thermal deformation processing-microstructure-formability relationship of the Ni_(47)Ti_(33)Hf_(20) shape memory alloy and can potentially be applied to other alloys.展开更多
文摘The study of how to “control forming and performance” during the thermal deformation of metal materials has always been a central theme in academic research, particularly in addressing the processing challenges associated with difficult-to-form alloys that possess unique functionalities. However, neither the currently commonly used phenomenological constitutive model, physical constitutive model, Dynamic Material Model (DMM) thermal processing theoretical model, and Ruano-Wadsworth-Sherby (R-W-S) deformation mechanism map model incorporating dislocation density nor the reported machine learning method has established a universal model that can achieve a quantitative description of the process-microstructure-formability of thermal processing. It is only possible first to use modeling research to obtain the law of thermal deformation behavior of alloys and then use the results of microscopic characterization to verify the theory. The research methods lack the characteristics of diagnosis and prediction optimization. This study proposes a machine learning framework for optimizing the random forest (RF) model based on a multivariate decision tree, including microstructure images and hot working process parameter information. It predicts the critical performance parameters, energy dissipation behavior, optimal processing window, and softening mechanism of ternary shape memory alloy Ni_(47)Ti_(33)Hf_(20) in the hot working process. This model has a certain universality. It enables coupled analysis of image information and process parameter data and introduces the calculation and ranking of feature importance, reflecting the applicability of feature values in model construction. Finally, the visualization technique Grad-CAM describes the correlation between the input microscopic image and the output, generating critical hotspots in the heat map. The model of accuracy in predicting the power dissipation rate is confirmed by the grain misorientation angles, thus realizing the establishment of a mechanism-driven model based on the evolution of critical microscopic structures during the thermal deformation of the alloy, which dramatically improves the interpretability of the machine learning model. This machine learning framework provides valuable guidance for quantitatively predicting the thermal deformation processing-microstructure-formability relationship of the Ni_(47)Ti_(33)Hf_(20) shape memory alloy and can potentially be applied to other alloys.
