Michael Frank Deering: AI: In-Process IC Inspection
Company: Fairchild Semiconductor (Schlumberger)
Years of Research: 1982-1985
At Shclumberger’s Fairchild Laboratory for AI Research (FLAIR) (later renamed SPAR), we had a number of AI computer vision people who liked to work on real (not pure research) problems, and an Integrated Circuit FAB line downstairs … so we started looking around to see how we might be able apply computer vision techniques to problems in manufacturing IC’s.
A lot of hard vision problems involve 3D, so looking at 2D chips sounded like a simpler start. After poking around a bit, we found that some parts of the quality control of IC fabrication was run on faith, or by having a number of low paid people looking through microscopes for defects in the constructed microcircuitry. So Michael Baird (home page: http://mikebaird.com/) just hired in from many years of applied vision work at GM, and I, started a project to use a computer to inspect in-process IC wafers.
I learned a lot more about types of IC production defects on this project than I ever needed again in years of IC design.
A dust particle landing on a IC wafer at a particular point in the fabrication might cause too much, or too little material to be deposited at a particular location in a circuit. Depending on the side of the dust particle and its location, this could be a non problem, a major short circuit, or a “mouse bite” out of a wire that will electrically test well but cause the circuit to prematurely fail somewhere down the road.
Even at the now relatively large FAB size of 2 microns transistor width, a scratch or a (human) hair on an IC wafer is a major slash across multiple parts of a circuit.
The first thing we had to do was to get a digital image of portions of an IC wafer into the computer. This involved a computer controlled microscope, computer controlled microscope stage, a video camera, and a high quality video frame grabber (my excuse to purchase an Ikons from Nick England and Mary Whitton). We also had to learn a lot about microscopy, micro-stage motion control hardware and software, and digitizer glitches.
I then wrote image processing software to: focus the microscope, find alignment marks on the wafer, find a particular die on the wafer, extract a digital image of the (visible) circuitry, and compare it to a known good image. This was not a simple pattern match, as IC process variation allowed the circuitry to change a lot and still be considered good.
The results was a research prototype system that was able to inspect chips that it was trained on with an error rate better than human operators.
Fairchild decided that they didn’t want to be in the business of commercializing the system, and neither at the time did the test and measurement division of Schlumberger (though this changed somewhat later). I had already gone on to VSP and FAIM1, and having made the prototype work, couldn’t see the point in continuing working on a project past the point where the executive decision to productize the system was on indefinite hold. Other companies and start-ups did productize similar products; and although I haven’t personally checked, I’ve been told that some non trivial portion of KLA/Tencor’s current business involves such products.
Besides internal SPAR technical reports, I don’t believe there were any external papers published on the project, although several public presentations were made. There was a demo video tape made, and may eventually turn up here in digital form.