Written by the Lenspathio engineering team. Process-node accuracy, false-positive methodology, SECS/GEM integration, and inspection economics.
Rule-based inspection systems use hand-tuned brightness thresholds and geometric criteria. CNNs learn feature representations from labeled defect images. The practical differences matter more than the algorithm headline.
Read articleA 3% false-positive rate means 3 wafers per hundred are sent for unnecessary review. At 120 wafers per hour, that's 3.6 lots unnecessarily escalated per 8-hour shift. The downstream costs accumulate quickly.
Read articleGEM E30 compliance isn't just a checkbox. Understanding the communication model — S6F11 event messages, collection events, variable reports — helps you write integration checklists that actually work in practice.
Read articleDetection rate variance between first and third shift is well documented in academic literature but rarely quantified inside production fabs. We analyzed 6 months of inspection data from a 300mm logic facility to put numbers to the pattern.
Read articleAt 28nm, a missed particle defect might affect one die. At 5nm, the same particle size can bridge critical features across multiple die on a reticle field. The consequence model for inspection errors changes fundamentally.
Read articleThis isn't a "machines replace humans" argument. It's a data question: under what conditions does automated optical inspection match or exceed human inspector performance, and what determines where it falls short?
Read article3–4 week evaluation on your process recipes and wafer sets. No commitment until you've seen the false-positive data.
