3D CNN and grad-CAM based visualization for predicting generation of dislocation clusters in multicrystalline silicon

Abstract

We propose a machine learning-based technique to address the crystallographic characteristics responsible for the generation of crystal defects. A convolutional neural network was trained with pairs of optical images that display the characteristics of the crystal and photoluminescence images that show the distributions of crystal defects. The model was trained to predict the existence of crystal defects at the center pixel of the given image from its optical features. Prediction accuracy and separability were enhanced by feeding three-dimensional data and data augmentation. The prediction was successful with a high area under the curve of over 0.9 in a receiver operating characteristic curve. Likelihood maps showing the distributions of the predicted defects are in good resemblance with the correct distributions. Using the trained model, we visualized the most important regions to the predicted class by gradient-based class activation mapping. The extracted regions were found to contain mostly particular grains where the grain boundaries changed greatly due to crystal growth and clusters of small grains. This technique is beneficial in providing a rapid and statistical analysis of various crystal characteristics because the features of optical images are often complex and difficult to interpret. The interpretations can help us understand the physics of crystal growth and the effects of crystallographic characteristics on the generation of detrimental defects. We believe that this technique will contribute to the development of a better fabrication process for high-performance multicrystalline materials.