CBAM-enhanced lightweight CNN for wafer map defect classification

Automated interpretation of wafer maps is central to manufacturing quality monitoring. Identifying rare defects with less detailed wafer maps is a challenging task. Moreover, class imbalance, heavyweight backbones, and limited model transparency are constraints for the real-world deployment of defective wafer identification. However, a nine-class wafer-map classifier is required that maintains high accuracy under tight parameter and compute budgets and provides decision-level interpretability, despite long-tailed class distributions. To address this issue, a compact convolutional network is presented for wafer-map classification on standardized low-resolution inputs. The architecture uses two convolution–pooling stages, followed by a modified convolutional block attention module (CBAM). Channel attention is realized via a shared multilayer perceptron with batch normalization for stable reweighting, while spatial attention uses a multi-scale gate to emphasize ring-like, edge-localized, and streak patterns. A compact dense head with softmax produces nine class probabilities, with a total footprint of approximately 0.15M parameters. Class imbalance is mitigated through a training-only convolutional autoencoder that generates minority samples via latent feature variation, together with a controlled reduction in the dominant None class. Validation and test sets remain unchanged. A fixed-seed protocol ensures reproducibility, and performance is evaluated using accuracy and macro-averaged precision, recall, and F1. On a balanced benchmark derived from the WM-811K dataset, the model achieves 99.88% test accuracy with near-ceiling macro-F1 while using a small fraction of the parameters required by transfer learning and transformer baselines and consistently outperforming conventional convolutional neural network (CNN) backbones. Post-training interpretability analyses with Grad-CAM, integrated gradients (IG), and occlusion show alignment between salient regions and physically meaningful defect morphology. Ablation studies indicate complementary gains from latent feature augmentation and attention mechanisms, while robustness checks with input noise and reduced training support show graceful degradation. The resulting pipeline is accurate, lightweight, and transparent, making it suitable for inline screening scenarios.