An Efficient Detection Method for Wafer-Test-Induced Defects

The wafer test is an essential component of semiconductor manufacturing. Rapid and accurate diagnosis of defects in a wafer test produces practical industrial values. However, the existing deep learning-based methods for wafer testing, although achieving excellent performance, face challenges in their reliance on extensive training datasets and the intricacy of models. To address these challenges, this paper proposes an efficient detection method based on inductive transfer learning for wafer-test-induced defects. By exploiting the visual feature extraction capability of the pre-trained model, the method requires only hundreds of wafer map data points for fine-tuning and achieves high detection accuracy. In addition, a progressive model-pruning flow is proposed to compress the model while maintaining accuracy. Experimental results show that the proposed method achieves detection accuracy as high as 100% for the wafer-test-induced defects on the validation set, while the pruning flow reduces the model size by 10.2% and the number of computational operations by 83.7%. The detection method achieves 61.7 FPS processing speed on an FPGA-based accelerator, enabling real-time defect detection.