Multimodal deep learning framework to predict strain localization of Mg/LPSO two-phase alloys

Abstract

This study proposes a method for predicting three-dimensional (3D) local strain distribution under compressive deformation of as-cast Mg/LPSO two-phase alloys from 3D microstructure images. The 3D local strain distribution was obtained by applying the digital volume correlation method to X-ray CT images before and after compression tests. Three microstructure descriptors were extracted from the 3D microstructure images around each strain measurement point: volume fractions of the phases, persistent diagrams that can express the connectivity of the phases, and two-phase spatial correlation that can express the spatial distribution of the phases. A deep learning model was then constructed to predict local strain from the three microstructure descriptors. Since two types of descriptors were used in this study, numerical data and image data, multimodal deep learning was employed to make predictions. Thus, the use of multiple microstructure descriptors enabled predictions to be made with higher accuracy than when predictions were made from a single descriptor. Feature importance of the descriptors was assessed through correlation analysis and occlusion sensitivity analysis. The results revealed that high strain tended to occur in the region where the hard phase, LPSO phase, had a large elongated phase oriented at a 45° direction to the loading direction. This result is consistent with other previous studies and indicates that the proposed method is effective in elucidating the relationship between the microstructure and the deformation behavior of the material.