Abstract
A considerable fraction of Czochralski silicon ingots undergoes remelting mainly because of dislocation generated during the growth, commonly termed structure loss. Being able to identify and categorize these failed ingots is a key step for understanding the root causes of structure loss and achieving a higher production yield. This work reports the utilization of machine learning (ML) to classify monocrystalline silicon ingots that have experienced structural loss during the Czochralski process. Three ML pipelines are implemented using different convolutional neural network architectures to analyze the surface images of the ingots. The accuracy and stability of the three ML pipelines are assessed. The results indicate that the pipeline that combines a pretrained model with an incremental training strategy obtains the highest accuracies and stable trainings of all tested pipelines, thereby making it the most suitable classification of structure loss in Czochralski-grown ingots.