Wafer composite defect recognition framework based on residual dynamic perception network with asymmetric multi-label loss

Advancements in semiconductor manufacturing require greater accuracy in wafer pattern recognition (DPR) to improve production yield. While current studies focus mainly on single-type defects, defects are more predominately mixed-patterned in actual industrial production scenarios. Moreover, the inherent noise in wafer maps, the inter-class similarity and intra-class discrepancy in defect categories further complicated DPR. To sidestep the difficulties, this paper proposes a novel deep-learning-based approach, namely the Residual Dynamic Perception Network (RDP-Net), for automating DPR. Its core innovation is the Dynamic Perception Mechanism (DPM), which adaptively fuses information from multiple convolutional kernels, layers, channels, and resolutions. DPM leverages cross-channel and pixelated perception generators to facilitate local-to-global feature extraction and fusion, effectively discarding irrelevant information that could lead to misinterpretation. Meanwhile, the Asymmetric Multi-Label Loss (ASL) is introduced, which helps balance the probability of positive and negative samples and mitigate the impact of incorrect labels. By combining the advantages of both methods, our RDP-Net reaches 99.13 % accuracy on the MixedWM38 dataset, outperforming current state-of-the-art DPR methods. In addition, visualized experiments clearly demonstrate the feature learning mechanism of DPM, showing significant interpretability. Further discussions covering the effect of different loss functions, the robustness to noise and mislabeling, the impact of inter-class similarity and intra-class discrepancy, as well as the required computation time, are also provided to demonstrate the feasibility of our proposed method.