Improved Wafer Defect Pattern Classification in Semiconductor Manufacturing Using Deep Learning and Explainable AI

In semiconductor manufacturing, the precision in identifying and classifying defect patterns on wafer maps is crucial for maintaining quality control and optimizing production yield. Traditional methods frequently suffer from the complexity and variability of defect patterns, necessitating more robust and automated solutions. This study introduces an improved deep learningSemiconductor ManufacturingDeep LearningDeep Learning framework utilizing a modified AlexNet architecture, referred to as Mod-AlexNet, specifically tailored for the classification of wafer map defects. The model leverages convolutional layers for automatic feature extractionFeature Extraction, ReLU activations for non-linear transformation, andBatch Normalization for improved training stability. The training process is optimized using the Adam algorithm with categorical cross-entropy loss, and incorporates early stopping and learning rate reduction to mitigate overfitting. To ensure interpretability, Local Interpretable Model-agnostic ExplanationsLIME (Local Interpretable Model-Agnostic Explanations) (LIME) and SHapley Additive exPlanations (SHAP) were employed. LIME provides local explanations by approximating the model with simpler, interpretable models to highlight crucial regions influencing predictions. SHAP offers a quantitative analysis by assigning each pixel a value representing its contribution to the final prediction, thus clarifying the model’s decision-making process. The proposed Mod-AlexNet achieved high classification accuracy, with LIME and SHAP effectively identifying key regions and quantifying feature contributions, respectively, thus confirming the model’s reliability and interpretability. This method offers an effective solution for wafer defect classification, addressing critical challenges in semiconductor manufacturingSemiconductor Manufacturing and enhancing the reliability of AI applications.