Classifying mixed-defect from single-defect training in imbalanced wafer maps via diffusion and attention

Accurate wafer bin map (WBM) classification in semiconductor manufacturing is crucial for identifying the root causes of defects and improving production efficiency. As process complexity increases, mixed-defect instances become more frequent in WBMs, highlighting the need for robust classification methods that can identify multiple defect origins. However, in real-world scenarios, single-defect WBMs are often class imbalanced, and mixed-defect WBMs lack labels due to high annotation costs, making classification particularly challenging. Therefore, we propose the first method that is capable of robustly classifying both single- and mixed-defect WBMs using only class-imbalanced single-defect training data. Our approach leverages a generalized zero-shot learning framework that classifies both seen and unseen classes by exploiting semantic information and employs a diffusion model as a generative classifier. The proposed method consists of two key components. First, a class-mixing diffusion model is designed to enhance the generation quality of underrepresented mixed and tail classes by learning from potential class combinations. Second, the classification performance achieved for seen classes is improved by using the prompt-wise attention score similarity to compare the generation processes for the training and test samples. Experiments conducted on the MixedWM38 dataset, which is commonly used in WBM classification research, demonstrate that our method achieves superior generalization performance with less overfitting to the training data and realizes significantly improved classification accuracy owing to its enhanced generation quality.