Mutlicrystalline Informatics—A Methodology to Realize High-Performance Multicrystalline Materials

We introduce our methodology, “multicrystalline informatics”, which integrates experiments, theory, computation, and machine learning to establish a universal guideline to maximize the macroscopic performance of multicrystalline materials by controlling microstructures and reducing crystal defects. We employ silicon as a model material and prepare various helpful machine learning tools for efficient materials development.<br/><br/>A realistic three-dimensional multicrystalline model of a silicon ingot for solar cells grown by directional solidification was obtained by integrating grain segmentation and orientation prediction by a machine learning model using multidimensional optical images of silicon wafers. The region of the interest was selected to contain the generation point of dislocation clusters as evidenced by photoluminescence images to reveal the spatial distribution of crystal defects. Finite element stress analysis was performed on the three-dimensional model by referring to the temperature distribution during crystal growth obtained from the simulation. The analysis results indicated that dislocations were generated along the main slip system where the greatest shear stress was applied. As for the experimental approach, multiscale structural analysis including transmission electron microscopy clarified that dislocations were generated from a curved Σ3 grain boundary with stepped microscopic {111} facets. It seems that high shear stress at the step edge would be responsible for the generation of dislocations.<br/><br/>As an alternative approach to access the underlying mechanism of the generation of dislocations, we attempt to apply image translation methods such as generative adversarial networks to obtain the likelihood distribution of dislocations generation from optical images or other processed images. By considering various models employing different images as a pseudo color input image, combining a grain boundary image and two optical images with different illumination angles showed a realistic output image. This indicates that a grain boundary network and crystal orientations play an essential role in the generation of dislocations. A further study to apply explainable artificial intelligence may help clarify the physics behind the phenomena.