Damage Analysis in Dual-Phase Steel Using Deep Learning: Transfer from Uniaxial to Biaxial Straining Conditions by Image Data Augmentation

Abstract

AbstractMicrostructural damage can occur during metal forming, but how and where this happens vary with the local microstructure and strain path. Large-scale analysis of such damage mechanisms is particularly important in advanced steels with a heterogeneous phase distribution. In our previous work, we demonstrated that deep learning enables a mechanism-based, statistical analysis by classifying many individual damage sites. The aim of this work is to generalize this approach to different stress states, e.g., biaxial instead of uniaxial tension, without manually labeling a large new ground-truth dataset of further micrographs and to thereby assess the changes in damage behavior with respect to stress state. Data augmentation and regularization allow us to directly apply our approach to the new, biaxial loading case. Overall, the network performance could be greatly improved and an analysis of changes in damage behavior, here the martensite crack angle distribution, with stress state can now be performed.