Latent-Space Defect Detection for Die-to-Die Wafer Variation

The patent introduces a learned similarity metric as an intermediary between unsupervised defect detection and supervised approaches. This metric, learned from data, acts as a mediator that enhances the detection sensitivity for targeted defect types while maintaining the unsupervised framework's ease of implementation. The similarity metric transforms the comparison between reference and test images, enabling selective sensitivity enhancement without requiring labeled training data.

The patent enables the system to learn the similarity metric automatically from the data without requiring manual labeling of defect candidates. The unsupervised learning approach allows the system to self-service by deriving the detection metric directly from the reference and test images, eliminating the time-consuming process of collecting and labeling training data while still achieving selective sensitivity enhancement.

The patent transforms the reference image generation process by changing from fixed statistical operations (median/average) to a learned transformation. The similarity metric is learned from the data, allowing it to adapt to die-to-die process variations. This parameter change enables the system to maintain simplicity in the generation process while significantly improving reliability in handling process variations through data-driven adaptation.

The patent applies local quality by enabling the learned similarity metric to operate at different spatial scales and adapt to local characteristics of sub-regions on the wafer. Rather than applying a global low-rank approximation that may destroy local defect signals, the learned metric can be tuned to preserve local features while still addressing die-to-die process variations. This allows different parts of the wafer to be handled with appropriate sensitivity.