Machine learning for early detection of structure loss in the Czochralski process

Abstract

This thesis investigates the use of machine learning techniques to predict structural loss in the Czochralski process. The Czochralski process is the industry standard for producing high-quality mono-crystalline silicon ingots for solar cells. The study evaluates the performance of three machine learning models; logistic regression, random forest, and neural network models across four regions of the ingot: neck, crown, shoulder, and body. The primary goal is to determine which model offers the best predictive performance for early detection of structural loss, thereby enhancing the efficiency and yield of the Czochralski process. The research reveals that the random forest model consistently delivers the highest accuracy, precision, and recall, especially in the neck and crown regions. This model effectively identifies the early signs of structural loss, making it a valuable tool for improving the process. However, all models faced difficulties in the shoulder and body regions, indicating the need for further refinement and more targeted features. Additionally, a time-saving model was developed to find time saved during the process by using the random forest model. By maintaining an accuracy threshold of 70%, this model achieved significant time savings, reducing the time required for remelting operations by 16 to 21.5 hours when tested on 51 ingots. These results highlight the potential of machine learning to enhance the Czochralski process, reducing production time and improving the quality of silicon ingots. Overall, the results demonstrate the potential for machine learning to significantly improve the Czochralski process, by enabling early detection of structural loss. Thereby, reducing the time required for remelting operations and enhancing ingot quality.