Mixed-defect Wafer Map Classification using CapsNet-based Models with Precise Scratch-pattern Reconstruction

Wafer maps of large-scale integration chips are matrices showing the defective chip position on a wafer. Determining defect patterns on wafer maps is crucial because these patterns indicate the root cause of defects in the manufacturing process. Convolutional neural networks can successfully extract local features; however, they cannot accurately identify linear scratch patterns widely spread on wafer maps, resulting in low classification accuracy. By contrast, CapsNet stores extracted features in multidimensional vectors (capsules) and learns the positional relationship between features using the capsule layers. Moreover, it learns input data features using extracted capsules by reducing the reconstruction errors of the decoder network. Mixed-defect wafer maps on a public dataset, Mixed-WM38, were classified using a CapsNet-based model. This model was developed by reconfiguring computational layers in the decoder and classifier networks to improve the reconstruction accuracy of scratch patterns. A synthesized high-resolution wafer map dataset and original pixel-wise metrics were used for scratch-pattern reconstruction experiments. The developed model could accurately recognize the fine feature of scratch patterns. Thus, it exhibited higher classification performance on the public dataset than benchmark models (F1: 98.1%). Wafer maps of real-world products were classified using the CapsNet-based model. Defect patterns, including scratches, were clearly identified, even in cases with multiple defect types overlapping on the same wafer map.