How an inspection-tool maker can let Micius check every chip image and leave only the uncertain ones to inspectors

On a back-end inspection line, an inspector sits beside each tool and looks at every chip image it takes, checking for chipped edges, chipped corners and cracks. A semiconductor equipment maker that builds these tools wanted Micius, Zhiman’s model, to make those calls instead, without changing its tools or the operators’ workflow.

The tool images one chip after another. For each image, the inspector traces the edge looking for a notch, a missing corner or a crack, logs the result, and decides whether the part passes or is held. By the time one image is done, the next is waiting. The tool only takes the picture; everything after that is a person’s job, so each tool runs only as fast as its inspector can look and decide. For the tool maker, each tool it sells means one more inspector at the customer’s site.

Checking by eye has two further problems. Hours of near-identical images wear people down, so misses and wrong calls can creep in as a shift goes on. And the standard lives in people’s heads: how big a notch counts as chipping, and whether it’s worth holding the part, can shift from one inspector or shift to the next.

The tool maker wanted Micius to take over that step and report several defect types against one standard. The conditions were set from the start: it had to run reliably on the tool, work on existing tools without changing the customer’s workflow, and repeat across many more tools and sites.

Micius works from two things the customer provides: a written standard for each defect type, covering how to recognise edge chipping, corner chipping and cracks and when a part passes or is held, and a good set of samples that includes good parts as well as defects. Good parts matter because defects alone only show how many defects get caught; with good parts you can also see how many get held by mistake. On the tool side, the software has to output images and accept results, and imaging conditions and resolution should be reasonably stable.

Micius has been added to the maker’s tools. By design, the tools, the imaging and the operators’ workflow stay as they are, and each image is checked on site, next to the tool, with no outside network.

For every image, Micius first decides whether the chip is good or defective. If it’s defective, it names the defect, edge chipping, corner chipping or a crack, and marks where it is. The result goes back to the tool software through a standard interface, and the tool software passes the part, holds it or sends it for review, by the customer’s rules. The results Micius is sure of go straight back; only the images it isn’t sure about wait for the inspector.

In one lot, images arrive one by one and Micius checks each in turn. Good parts go straight back to the tool software. An edge chip goes back with its location, and the tool holds the part. One image looks as if the upper-right corner is missing a piece, but glare hides the detail, so Micius sends it to the inspector instead of deciding. When the lot finishes, that is the only image the inspector has to look at.

The decision is the inspector’s: mark the part good, or confirm the defect and hold it. The standard behind that decision stays the customer’s too. As the customer supplies samples of new defects, new types can be added without rebuilding the inspection system.

Inspection has to avoid two failures: escapes, where bad parts get through, and overkill, where good parts are held. The validation measured both, plus speed. On test samples supplied by the customer, Micius’s calls on edge and corner chipping were checked for catch rate and speed. In both test sets, it caught more than 98% of defective images and wrongly flagged about 0.3% of good parts. Each image took about 0.1 seconds. Every image in testing was checked automatically, with no human review. That overkill rate works out to about 3 wrongly flagged images in every 1,000 good parts.

These are test-sample results, not production acceptance figures. They cover edge and corner chipping only, not cracks or other defects, and each new site has to confirm them on its own images, standards and acceptance data. Catch-rate, overkill and speed targets should be agreed before that testing starts.

For the tool maker, checking images with Micius can become a feature that ships with the tool. Once the model, the service and the interface are standardised, it can go out to more tools and customer sites.