What does the In-Sight 1750 do?

Cognex has launched the In-Sight 1750 Series for semiconductor manufacturers. The system combines AI-based reading, advanced imaging, and dedicated illumination to decode identification marks on wafers and panels. Its target is the traceability demands of high-volume fabs, where each carrier and substrate must remain identifiable through successive process tools.

Wafer identification happens at automated handoffs inside the cleanroom — load ports, stockers, conveyor stations. The marks themselves are application-specific, and so are the substrates. Modern fabs run polished silicon, transparent carriers, bonded stacks, and panel formats for advanced packaging. The reading problem is that optical contrast between mark and surface depends on the substrate, so a reader tuned for one combination underperforms on another.

Cognex's answer is to combine an AI-trained decoder with a matched optical stack — sensor, lens, and illumination built as one unit rather than a configurable parts list. The vendor's stated positioning, as reported by eeNews Europe, is that the system "improves traceability while keeping production lines moving," with the combination of AI, imaging, and illumination aimed at the "difficult wafer markings and challenging substrate materials" that conventional tools handle poorly.

What does AI change in the reading chain?

The physics is straightforward. Optical mark contrast depends on the reflectance difference between the mark and the surrounding surface. Illumination wavelength, angle, and polarization control that difference. Polished silicon returns a strong specular component; rough or low-reflectivity coatings return more diffuse light and add subsurface scatter, which confuses a rule-based reader looking for clean edges and quiet zones.

AI enters where the contrast gets too low or too variable for rule-based OCR. A trained model can localize a degraded mark against a noisy background, then pass the cropped image to a decoder. The benefit is not that the camera sees better, but that the algorithm tolerates more variation in mark quality and substrate. That tolerance is what the vendor is selling.

What remains to be qualified before adoption?

That tolerance matters because a no-read at a single handoff stops the lot. The cost of a stop is the entire downstream tool group's idle time, not just the misread wafer. Buyers will therefore want:

• Read-rate numbers on their own substrate mix
• False-pass rate when a damaged mark is misread
• Calibration and retraining interval when a new product enters the line
• Recovery behavior when the AI model's confidence falls below threshold

Cognex has not published those figures. The evaluation therefore moves from the press release to the load port.

Compliance is the other open question. Wafer and carrier identification feeds the traceability records chipmakers maintain for yield analysis, lot genealogy, and audit purposes. Any reader that misreads a lot identifier propagates a data-integrity error through every downstream process step.

Fabs must therefore qualify the 1750 against the same acceptance criteria they apply to existing readers — not just on read rate but on the false-read rate that determines how often human review must step in. That qualification, not the launch announcement, decides whether the system earns its place above the load port.