Automated optical inspection in semiconductor fabs produces up to 50% false positive defect classifications, forcing human operators to manually re-review thousands of images per shift and delaying yield-learning cycles

Semiconductor fabrication facilities rely on Automated Optical Inspection (AOI) systems to detect wafer defects after critical process steps like lithography, etch, and deposition. However, rule-based AOI systems generate false positive rates as high as 50%, flagging normal process variation -- such as grain boundaries, surface texture changes, and measurement artifacts -- as defects. Each false alarm requires a human operator to pull up the high-resolution SEM image, classify the defect, and disposition the wafer, consuming hours of skilled labor per shift. This false-alarm burden delays the yield-learning feedback loop that is essential for ramping new process nodes to volume production. Why it matters: engineers spend more time reviewing false alarms than analyzing real defects, so the time from defect occurrence to root-cause identification stretches from hours to days, so yield improvement during new node ramps slows by weeks, so fabs lose millions of dollars in foregone good-die revenue during the extended ramp period, so the entire semiconductor industry's ability to deliver next-generation chips on schedule depends partly on solving a classification accuracy problem in optical inspection tools. The structural root cause is that rule-based AOI classification relies on static threshold parameters (brightness, contrast, size) that cannot adapt to the inherent process variability across a wafer and between wafers, and the defect types that matter most at advanced nodes -- subtle pattern distortions like line-edge roughness, microbridging, and overlay errors -- have visual signatures that overlap heavily with normal process variation, making binary threshold-based classification fundamentally inadequate.