Scalable and lightweight vision transformers for mixed-defect classification in semiconductor wafer maps

Semiconductor wafer defect classification faces critical challenges from overlapping defect patterns, severe class imbalance, and deployment constraints in resource-limited manufacturing environments. We propose ViT-Tiny, a lightweight Vision Transformer architecture optimised for mixed-defect classification with only 5.7M parameters — $$15\times$$ fewer than standard Vision Transformers. Through domain-specific adaptations including class-balanced loss and defect-aware augmentation, ViT-Tiny captures both local detail and global context via $$16\times 16$$ patch-based self-attention. On the WM-38K wafer map dataset our model achieves 98.4% accuracy, outperforming larger transformers (Swin, MSF-Trans) and deep CNNs (ResNet, EfficientNet) by up to 2.8% in precision while excelling on complex 2-way, 3-way, and 4-way mixed defects. The proposed model has been validated on MVTec-AD demonstrates strong generalization with a 13.75% average accuracy improvement over baselines. With $$1.77\times$$ faster inference and $$4.47\times$$ lower memory footprint than comparable transformers, ViT-Tiny provides an efficient, scalable solution for real-time industrial defect inspection.