Defect Prediction and Optimization in Semiconductor Manufacturing Using Explainable AutoML

The semiconductor manufacturing industry often faces the severe challenges of data scarcity and imbalance. While the semiconductor industry has conducted extensive research on leveraging machine learning to improve yield, defect prediction remains largely unexplored, especially with small datasets. This research proposes a framework called xAutoML, which automatically selects the optimal model and hyperparameters for defect prediction to enhance the interpretability of the results. Furthermore, it addresses the critical issue of data imbalance, a common problem in defect prediction tasks, by employing techniques such as focus loss and oversampling. We use publicly available datasets to demonstrate how xAutoML effectively adapts to data constraints and deeply analyzes key features influencing defect occurrence. Results show that the proposed method outperforms traditional methods in terms of prediction accuracy and the provision of actionable and interpretable insights. Its application in real-time defect monitoring and process optimization in semiconductor manufacturing helps bridge the gap between advanced machine learning techniques and practical industry applications.