Digital imaging comes to second-optical IC inspection

Edina, MN--The human eye responds automatically: the pupil dilates to adjust for light, and refocuses the image. That's the kind of automated control that August Technologies' new NSX-80 system uses for second-optical inspection of semiconductors. The system employs high-resolution digital imaging technology and a proprietary algorithm that allows it to "learn" to distinguish good wafers from bad. August's President Jeff O'Dell claims the system operates five to 10 times faster than human inspectors, and is also more accurate.

Growing up alongside technologies such as machine vision, reasonably priced charged-coupled-device (CCD) array sensors, and explosive growth in semiconductors, August's focus is on applying machine vision to the semiconductor industry. Semiconductor production has developed with widely recognized stages of inspection, commonly referred to as first, second, and third optical inspection. While first and third inspections were automated early on, the second has long relied on the human factor.

Until now, second inspection has been a labor-intensive process that took advantage of the fact that the human eye and brain were the best available technology, explains O'Dell. "Operators sit at microscopes, hour after hour, visually scanning wafers and looking for anomalies. Computers just are not very good at recognizing anomalies. You have to sort out what's different yet acceptable, from what's different and not acceptable."

But the human-inspection machine isn't perfect either, says O'Dell. People recognize the anomalies easily at first, but they slow down over the course of a day. They make errors as the hours creep by. Soon semiconductor wafers all look the same. The NSX-80 solves this problem. It employs a combination of newly available sensor technology and computer technology, and a proprietary algorithm to help the system "learn" what is and isn't an acceptable semiconductor.

To capture the wafer's image digitally, the system uses a Kodak MegaPlus camera. Its million-pixel CCD array allows the NSX-80 system to resolve anomalies down to 0.5m in size. The camera is controlled through an RS-422 interface to a Matrox graphics card on a Pentium-based Windows NT platform. It captures from eight to 15 frames/sec, with exposures as short as 15 ms. "The MegaPlus camera gives us the resolution we require with a very acceptable price/performance point," O'Dell says. "That value wasn't available until quite recently. It's also highly configurable and controllable, just like the human eye."

While Kodak's camera technology replaces human eyes, a second question remains. How do you develop an algorithm that reliably emulates the way a human mind would "learn" this process? Images are captured and processed in the computer using a sophisticated image processing algorithm. "This algorithm has been the most difficult piece of the puzzle," O'Dell says. The first optical inspection takes place when the semiconductor is less complex. But second optical takes place after the wafer is complete and has been electrically tested, a process that may leave marks on it.

The inspection system must distinguish between these random markings and a faulty wafer. "This involves developing a database of acceptable images and unacceptable images, and recognizing the characteristics that distinguish them," O'Dell says. "The modeling is incredibly complex." The result is a system that can inspect five to 10 times faster than a human inspector. And with greater accuracy.

Give the same human inspector the same wafer at two different times and he or she may come up with different answers, O'Dell notes. Give the same wafer to different inspectors and again, they may provide different answers. The bottom line is that people are subjective. Computers remove subjectivity from the inspection process, according to O'Dell.

The delay in automating the process has been a question of accuracy. Early algorithms might have caught only 90% of faulty semiconductors; or, an algorithm designed to be more rigorous, might identify 10 good wafers as defective. Such an error in either direction costs the manufacturer. O'Dell believes his company's new system cuts the error rate to under 1%.

The system can help manufacturers in other ways, too. Beyond simply identifying defective wafers, the system can help to identify where recurring defects are and what they look like. "Then it's up to the manufacturers to determine how to make this process information work for them," O'Dell says.