Open-world scanning electron microscope nano-defect category discovery via self-normal prototypes

Nanoscale defect classification is crucial for yield enhancement in integrated circuit (IC) manufacturing. However, existing methods rely heavily on static external datasets. Consequently, they struggle to adapt to the rapidly evolving layouts and novel defect categories found in open-world scenarios. We propose SNAP-SEM to tackle the dynamic variations of Scanning Electron Microscope nano-defects in open-world IC manufacturing. It is a novel nano-defect category discovery framework driven by Self-Normal Augmented Prototypes. First, a Self-Normal Prototype Extractor is designed to distill normal prototypes from within a single image, thereby breaking the reliance on external support sets and generating a robust spatial prior for defects. We then inject this spatial prior into the network’s attention mechanism. This integration encourages the model to suppress structured background interference and focus its representational capacity on localized anomalies, while preserving essential context. Finally, within a unified hyperspherical space, spatial prior-augmented category discovery is achieved via prototype learning, accomplishing the autonomous clustering and discovery of unknown defect categories. We construct and will open-source the IC-SEM dataset, which encompasses 13 defect types across three distinct manufacturing stages. Extensive experiments demonstrate the superior robustness and accuracy of SNAP-SEM on generalized category discovery tasks across diverse manufacturing stages. Furthermore, the proposed framework achieves an efficient inference speed of 182 ms per image and has been successfully deployed in actual IC manufacturing pipelines. Future research will explore integrating cross-modal layout data to address extreme defect scales.