Explainable machine learning for predicting the mechanical properties in bainitic steels

Microstructures in steel can be understood as hierarchical structures holding information on various length scales (e.g. from nano to macro scale). In electron microscopy images, the microscope resolution allows extraction of meso-scale information of grain interiors and boundaries encrypted in morphological features. So far, experts design experiments to extract and select important microstructural features based on their experience and evidence from literature. These features often serve as indicator for interpretation of corresponding mechanical properties. But such approaches commonly capture only a few features while others remain hidden in the microstructure. This becomes evident for complex bainitic microstructures in contrast to other microstructures in steel. In this study, we use a machine learning approach to correlate the most important morphological features, retrieved from segmented Martensite-Austenite (M−A) islands in microscopy images of bainite, with the impact energy from mechanical testing. The used open-source framework selects the most important features and explains their global and individual contributions on the model output. We test the learned correlation on generalizability of these microstructure-property linkages on bainitic steels with different chemical compositions and applied cooling regimes.