The single or mixed defects in wafer maps reflect critical problems in semiconductor manufacturing processes, thus their accurate recognition plays a pivotal role in root cause analysis of anomalies and process stability maintenance. The increasing complexity of mixed-type defects poses new challenges to the feature extraction capability and learning capability of current vision models. To address this challenge, we propose WM-EVA-ViT: a transferred pre-trained large-scale vision foundation model with supervised contrastive learning (SCL)-assisted fine-tuning for wafer map mixed defect recognition (WMMDR). The vision foundation model demonstrates accelerated learning capabilities during the fine-tuning process for defect feature extraction, leveraging its superior general visual feature extraction capacities. Furthermore, a SCL-assisted fine-tuning method is proposed, which enhances class-specific feature discrimination through contrastive learning with class label informed constraints. Experimental results on a real-world dataset validate the effectiveness and superiority of the proposed method. Besides, this method offers novel perspectives for WMMDR in the era of large-scale models.

wafer map,mixed defect,large-scale vision foundation model,supervised contrastive learning (SCL)