Abstract
This work investigates the relationships between process parameters and the occurrence of structure loss in Czochralski silicon ingots using machine learning. Subsets of features are identified from a dataset of over 14,000 ingots and are used to train random forests to predict structure loss with high accuracy. Multiple rounds of feature importance analysis and refinement are conducted to isolate the process parameters that may have the most significant impact in the occurrence of structure loss. Partial dependence analysis is employed to examine how variations in particular parameters might affect the likelihood of structure loss happening. The results show that the most predictive features of structure loss are primarily recorded late in the process. These features are often influenced by manual interventions or reflect the outcome of structure loss itself. In contrast, early-stage parameters exhibit limited predictive power, suggesting that either early indicators of structure loss are not captured in the available data or that structure loss originates from events occurring later in the growth process. While not predictive in a preventive sense, the model effectively detects deviations from normal operation, thereby demonstrating the value of machine learning for uncovering complex patterns in manufacturing processes data.