Enhancing Confidence and Interpretability of a CNN-Based Wafer Defect Classification Model Using Temperature Scaling and LIME

Accurate classification of defects in the semiconductor manufacturing process is critical for improving yield and ensuring quality. While previous works have mainly focused on improving classification accuracy, we propose a model that can simultaneously assess accuracy, prediction confidence, and interpretability in wafer defect classification. To solve the class imbalance problem, we used a weighted cross-entropy loss function and convolutional neural network–based model to achieve a high accuracy of 97.8% on the test dataset and applied a temperature-scaling technique to enhance confidence. Furthermore, by simultaneously employing local interpretable model-agnostic explanations and gradient-weighted class activation mapping, the rationale for the predictions of the model was visualized, allowing users to understand the decision-making process of the model from various perspectives. This research can provide a direction for the next generation of intelligent quality management systems by enhancing the applicability of the proposed model in actual semiconductor production sites through explainable predictions.