A machine learning approach to reveal microstructures contributing to dislocation clusters in multicrystalline silicon

We report on our attempts to leverage a machine learning model to investigate the crystallographic features contributing to the dislocation clusters in multicrystalline silicon. The developed model predicts the presence of dislocation clusters at the center of optical images and enables us to obtain their spatial distribution in a wafer by scanning the position for prediction. Information about the crystal orientation distribution was incorporated into the model by using multiple optical images of the textured wafers taken under illumination from various angles of incidence. The area under the curve, an indicator of prediction accuracy, of the optimized model was over 0.99, which is much higher than that of the previous model using optical images taken at a single illumination angle. In addition, the regions of interest in the predictions were evaluated using gradient-based class activation mapping. The results were analyzed using crystal orientation information, and we revealed crystallographic features of the regions surrounding the dislocation clusters. We then proposed a design guideline to produce multicrystalline silicon wafers with a low density of dislocation clusters based on our findings. The prediction of dislocation cluster distribution in the virtual multicrystalline silicon wafer following the guideline indicated that the number of dislocation clusters could be much lower than that in the conventional real multicrystalline silicon wafers. Accordingly, the technique proposed in this study should be effective for rapidly analyzing the relationship between the defects and the microstructure in multicrystalline materials and understanding the physics behind microstructure and defect evolution in the fabrication processes.