U-net segmentation of structured defect patterns in unpatterned wafer inspection maps using synthetic rule-based data

Abstract A U-shaped convolutional network (U-net)-based approach is proposed for segmenting spatially structured defects in unpatterned wafer inspection maps. The model is trained entirely on programmatically generated, rule-based synthetic data, without any real wafer inspection input. The synthetic dataset comprises both random point defects and structured patterns (e.g. linear and curved traces) represented as continuous or discontinuous sequences. This method addresses the limitations of conventional rule-based inspection techniques, which often fail to detect spatially coherent defect formations across wafer surfaces. After training, the U-net model effectively extracts these structured patterns and successfully highlights linear and arc-shaped defect formations using real inspection data. The findings demonstrate that spatial defect structures can be identified without the manual annotation of real wafer inspection data. Overall, the proposed framework offers a practical approach to automated defect segmentation in unpatterned wafers and serves as a valuable visual tool for process monitoring and failure analysis.