By David Lammers

As more (and better) data is coming from sensors on semiconductor production equipment, fab managers are evolving from reactive to predictive approaches to maintenance, scheduling, and yield improvements.

After several years of development, including pilot studies coordinated by Sematech/ISMI, chip makers are gradually evolving towards using virtual metrology, predictive maintenance, and related techniques, speakers said at last week’s 2011 ISMI Symposium, which attracted about 400 attendees to Austin, Texas.

Semiconductor companies have used tool data widely for fault detection and classification (FDC) and run-to-run (R2R) control. A few device makers are starting to use predictive maintenance, which relies on modeling of past failures, to suggest when expensive sub-components, such as an e-chuck, may need replacement. Predictive maintenance (PdM) can save fabs a million dollars or more per year by extending schedules for parts replacement, said John Scoville, senior director of applications engineering at the Applied Global Services operation of Applied Materials, who delivered a keynote speech at the symposium.

Recently, Applied announced that Toshiba has adopted Applied’s “SmartSched” predictive scheduling software, which Scoville said can improve the efficiency of the expensive lithography tools in a fab.

The “play forward” capabilities which build on virtual metrology and predictive maintenance will be extended to predict a fab’s yields and throughput. By knowing the likely health of the fab’s toolset for the near future, fab managers will be able to predict yields and estimate how many wafer starts will be needed to meet customers’ orders.

“Each one of these techniques builds on the others, progressing in parallel,” Scoville said. “The better we get at virtual metrology, predictive maintenance, and others, the more we can implement predictive yield applications.”

One scenario is that a foundry manager could choose to sacrifice yields a bit while boosting wafer throughput in order to meet a customer’s order and maximize the fab’s revenues. “We want to get to a point where fab managers are not bumping around in the dark. We want to know and predict the impact of decisions on factory performance,” he said. The vision is to have the predictive capabilities becoming more accurate “in lock step with reality,” with the models continually updated.

“Prediction can be a unifying capability, building on the automated data mining of terabytes of data,” Scoville said, adding that while today’s models tend to be “recipe specific” they need to become “more generic.”

David Stark, who manages the Sematech/ISMI program on predictive techniques, (which Sematech calls Prognostics and Health Management, or PHM) said PHM eventually will provide fab managers with a dashboard that will improve asset utilization. Now, a foundry’s managers may schedule WIP (work in process) based on the best possible human decision-making. With “tons of data” being extracted from tools, PHM will empower fab managers to make better-informed decisions.

“The WIP (chosen by humans) may not be the fab’s best WIP. The fab managers are doing their planning for the coming week, when to take tools down and so on. With a PHM system, those decisions would be less emotional.”

Stark said it took many years for the wider industry to implement FDC, for example. After two separate trials, involving two different fabs and two tool makers (Novellus and Lam Research), the PHM approach is now being put into commercial use.

While participants at the symposium said Applied has launched a wide-ranging initiative to gain synergies between its tools and the predictive techniques, other companies are also developing predictive techniques and solutions.

Tamara Byrne, an applications manager at MKS Instruments Umetrics, said the best predictive techniques draw upon the multivariate analysis techniques pioneered for FDC. Byrne, who earlier led IBM’s pioneering multivariate fault detection effort in Fishkill, N.Y., said process engineers in the pharmaceutical, paper, chemical, and semiconductor industries are gradually moving into virtual metrology and predictive techniques. Some of them buy the Umetrics Simca tool to analyze multivariate data, as part of in-house VM and PdM solutions.

“There are more in-house efforts underway than there are companies relying on external suppliers,” she said, adding that the large Asian fabs in particular want to hold techniques close to their vests. “Our focus is not just on semis, but I would say that chip companies do see big benefits from looking at the data – there is a big ROI to multivariate analysis.”

MKS Instruments Umetrics, with 60 employees worldwide including 40 developers in Sweden, doesn’t always know exactly what its customers are doing with its tools. “Process engineers like to be autonomous, and that can be a good thing, because you want a thinking person making the decisions. You never want the software to be a substitute for a good engineer,” Byrne said at the ISMI event.

Tom Ho, president of the U.S. operations of BISTel, a South Korea-based provider of an equipment engineering system (EES), said semiconductor vendors are looking for partners which can help them analyze data coming off the tools. BISTel, which counts Samsung as a key customer, is hiring Ph.D.-level scientists to create algorithms to mine data and develop FDC and PdM solutions. The company currently employs 170 people, most of them based in Korea.

“Our YiEES tool is monitoring the pressure, temperature and other variables. That results in lots of data, and we can help create the models, the algorithms, and to identify cause and effect relationships,” Ho said, adding that his company often complements “internal solutions.”

Many of the presentations at the two-day symposium, as well as the workshops and short courses, focused on advanced process control, tool monitoring, data mining, predictive techniques, virtual metrology, and related subjects.

Michael Mooney, of Edwards, described an effort to use fault detection data for predictive diagnostics, or model-based prognostics. The Edwards effort involves probability density functions (PDF). “We have found that neural networks are not good; rule-based approaches are better” because they provide better access to the models. The goal is to develop abatement equipment which “automatically learns about failure modes,” Mooney said.

Tags: Applied Materials, BISTel, Edwards, FDC, MKS Instruments, predictive maintenance, SmartSched, virtual metrology

This entry was posted on Monday, October 24th, 2011 at 9:50 pm and is filed under News Stories. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.