Nano Sentinel: An Explainable Deep Learning Framework for Semiconductor Wafer Defect Classification and Inspection Workflow Support

The wafer-map defect pattern classification is an important task in semiconductor manufacturing, as the spatial arrangement of defective dies can provide valuable information for identifying recurring failure patterns. The manual inspection of wafer maps can be extremely labor-intensive, especially if large volumes of inspection data have to be reviewed. Deep-learning methods provide a way to automate the pattern classification, but the classification predictions are only a part of an inspection workflow and do not present evidence for individual predictions. This paper introduces Nano Sentinel, a prototype framework that integrates deep-learning-based wafer-map classification with explainability and inspection workflow elements. The classification component uses MobileNetV2 transfer learning to classify wafer maps from the WM-811K dataset into nine Center, Donut, Edge-Loc, Edge-Ring, Loc, Near-full, Random, Scratch, and none classes. The framework incorporates Gradient-weighted Class Activation Mapping for model visual inspection, maximum softmax probability for confidence reporting, a provisional defect-severity assessment, rule-based inspection recommendations, and inspection record generation. A lot-aware partition was used for training and validation of the model. Preliminary experiments with the selected V2 model achieved 94.56% validation accuracy and 0.8332 macro-averaged F1-score. An independent diagnostic evaluation on the official test partition returned 88.66% accuracy and 0.4727 macro-F1, which suggests significant variance between the performance of different classes and highlights the need for further generalization studies. The workflow components were demonstrated as prototype decision support features, rather than a fully validated industrial inspection system. The study emphasized the combination of classification, visualization, and documentation elements while class imbalance, minority-class mistakes, and the validation of severity and rule-based decisions are important areas of future work.