Search

Accelerating design and development of new materials by establishing process-structure-property (PSP) linkages is one of the core contents of materials science. One of the challenges is how to accurately forecast the property by the features including chemical compositions, experiment conditions, and structure information. In this study, with consistent features in statistics and materials science, we proposed a microstructure-informatic strategy to achieve the goal of accurately predicting Vickers hardness of austenitic steels. Feature engineering including correlations analysis, importance ranking and microstructural features extraction was employed to ensure the most information contained in the features related to the property. Through training and comparing six regression models with different input features, we demonstrated that one of the models inputting microstructural features obtained by two-point statistics combined with principal component analysis (PCA) maintains the highest accuracy (absolute error≤13.63 MPa, relative error≤8.86%) and predictive stability (minimum error range). The excellent generalization ability of this model was validated by eight experimental instances unseen in the original dataset. We believe that our strategy can be used to guide future experiments due to its high precision. Most importantly, the strategy can be generalized to predict other mechanical properties controlled by microstructures in more material systems.

Accelerating design and development of new materials by establishing process-structure-property (PSP) linkages is one of the core contents of materials science. One of the challenges is how to accurately forecast the property by the features including chemical compositions, experiment conditions, and structure information. In this study, with consistent features in statistics and materials science, we proposed a microstructure-informatic strategy to achieve the goal of accurately predicting Vickers hardness of austenitic steels. Feature engineering including correlations analysis, importance ranking and microstructural features extraction was employed to ensure the most information contained in the features related to the property. Through training and comparing six regression models with different input features, we demonstrated that one of the models inputting microstructural features obtained by two-point statistics combined with principal component analysis (PCA) maintains the highest accuracy (absolute error≤13.63 MPa, relative error≤8.86%) and predictive stability (minimum error range). The excellent generalization ability of this model was validated by eight experimental instances unseen in the original dataset. We believe that our strategy can be used to guide future experiments due to its high precision. Most importantly, the strategy can be generalized to predict other mechanical properties controlled by microstructures in more material systems. Unlabelled Image • A microstructure-informatic strategy to accurately predict mechanical property of steels. • Integrated module by advanced techniques and approaches to adaptively capture the optimal mapping from input to output. • Prediction with small uncertainty of Vickers hardness of steels based on experiment data. • Informative microstructure features extracted by two-point statistics and principal component analysis methods. [ABSTRACT FROM AUTHOR]