Conditional Generation of Overlapping Mixed-type Wafer Bin Map Defects via Feature-wise Linear Modulation

<title>Abstract</title> <p>In semiconductor manufacturing, mixed-type wafer bin map (WBM) defects contain spatial evidence valuable for process diagnosis and yield management. Existing approaches, however, either provide image-level labels without localizing constituent defects or produce ambiguous predictions in overlapping regions, limiting their usefulness for engineering decision support. To address this gap, we propose the Wafer Conditional Generative Network (WCGN), a conditional generation framework that integrates Feature-wise Linear Modulation into a U-Net architecture. By conditioning the decoder on a target defect label, WCGN converts mixed-type WBM analysis into a label-conditioned mask generation task and disentangles concealed constituent patterns within overlapped defects. Experiments on MixedWM38 for classification and on a generated mixed-type benchmark with pixel-level masks for segmentation show that WCGN outperforms state-of-the-art baselines, achieving an Exact Match Ratio of 97.85\% and a mean Intersection over Union of 0.8725. The model also improves segmentation of thin scratch patterns and maintains strong performance on unseen defect combinations excluded from training. These results show that conditional mask generation can externalize process-relevant spatial evidence from complex WBMs rather than only output black-box labels. From an engineering informatics perspective, WCGN provides an interpretable representation of overlapping defect structures that can support engineer-in-the-loop diagnosis, process monitoring, and yield management in semiconductor manufacturing.</p>