Towards accurate prediction for ultra-low carbon tempered martensite property through the cross-correlated substructures.

[Display omitted] • A high-throughput experiment for steels based on LMD and GTHT was developed. A machine learning model possessing microstructure input terminals was put forward. • The average error of validation set in microhardness can be as low as 14.37 HV. • The reported strategy holds the promise for the application to other martensite steels. Accurately predicting properties of steels containing martensite by using models based on traditional strengthening mechanisms remains a challenge. In this study, a smart machine learning model possessing two-dimensional microstructure input terminals was developed using high-throughput experiments and machine learning on steels for low-temperature service. An algorithm based on a convolutional neural network enriched with the two-dimensional input terminals increased the prediction accuracy, achieving an average microhardness error of as low as 14.37 HV for the validation set. The improved prediction accuracy is ascribed to the comprehensive strengthening mechanism and coupling of strengthening effects contained in the multifarious input terminals. The information acquisition and cross-correlation of substructures related to strengthening mechanism played an important role. The reported strategy can deepen the cognition of the strengthening mechanism of tempered martensite. It is promising for application to different steels containing tempered martensite. [ABSTRACT FROM AUTHOR]