Ab Fab – Edwards turns to Hadoop in push for smarter semiconductor

In Industry 4.0 terms, semiconductor fabrication plants (or ‘fabs’) are among some of the smartest factories in the world.

The companies that own them have invested significantly in sensors and connectivity to monitor inventory levels, for example, and to spot manufacturing irregularities that could lead to problems with quality or yield. But there are still areas in which they could be getting more benefit from Internet of Things (IoT) technologies.

One potential area is the vacuum and abatement systems on which these factories rely for safe and clean operations. Vacuum systems keep cleanroom environments meticulously free of contaminants, while abatement systems act as exhausts, funnelling dangerous waste products away. In some cases, integrated units perform both functions.

Either way, around six out of ten semiconductor fabs worldwide buy vacuum and abatement equipment from a UK-headquartered specialist called Edwards, which has been part of Swedish industrial company Atlas Copco since 2014.

Edwards recently announced that it is using MapR’s distribution of the Hadoop big data framework to develop a platform that will analyse data from its equipment located at customer sites worldwide and enable the company to provide them with real-time anomaly detection and predictive maintenance services. As David Hacker, Edwards’ strategic marketing manager explained:

There’s an awful lot of metrology equipment that goes into fabs already today, but what we’re aiming to do is extend the reach of metrology. The data that’s available to us from our systems holds valuable clues as to what’s happening inside the process chambers where semiconductors are made, but that data’s largely been an untapped resource.

Putting data to work

In other words, there’s a big opportunity here for continuous monitoring. Looking at data relating to how an integrated vacuum/abatement system responds to different gas mixtures in the process chamber, for example, might provide a more reliable sign that a chemical reaction had reached its conclusion than established methods.

Hadoop makes good sense for this kind of project, says Hacker, because of the sheer volume of data involved. Even a medium-sized semiconductor plant could have as many as 4,000 pieces of Edwards equipment inside. A big fab might contain 8,000 or more.

That’s a lot of assets, producing a lot of data. Our older equipment produces data pretty slowly, but as we’ve updated our portfolio, we’ve got a lot of equipment now that offers much faster data rates, measured in milliseconds. So it was pretty clear some years back that our older SQL-based approach was running out of steam and that we needed a system with a high ingestion rate, which is where Hadoop-type systems come into play.

Edwards is working on this deployment with researchers from Fraunhofer-Gesellschaft (Fraunhofer Society), Europe’s largest institution for applied research. As Hacker explains:

The idea of the Fraunhofer project was that they were looking for a partner to work on Industry 4.0, and we thought that Industry 4.0, from a theoretical point of view, presented some quite nice solutions to bringing discipline into things we had wanted to do for quite some time around sensorization, around security, around cloud.

There are several strands to this partnership, but perhaps the most important is research into the correlation between manufacturing process events and pump behaviour. Machine learning techniques will be used to detect anomalies in sensor data, with the aim of developing and refining approaches to predictive maintenance. The MapR platform will be piloted in the semiconductor cleanroom environment at Fraunhofer EMFT, a part of the institution focused on research into sensors and actuators.

Some of the data fed into MapR will come from sensors native to Edwards’ equipment. This already provides around 20 parameters today with which researchers can work, according to Hacker. But as the project progresses, data from sensors external to the equipment and based elsewhere in cleanrooms will start to be introduced.

The ultimate goal, explains Hacker, is to commercialize this application of the MapR platform, selling it as a service to Edwards customers in the semiconductor industry. There are plenty of commercial details still to be worked out here, he says, but the company is working closely with a team from MapR to co-develop the solution. With the MapR framework now up and running, it’s time to get started on exploring the data, he says.

We’re at a really interesting and exciting point now. From now onwards, data will be flowing into MapR and we’ll be starting to explore the possibilities.

Image credit – Production and cleanroom facilities at work in Intel’s D1D/D1X plant in Hillsboro, Oregon, in April 2017. (Credit: Intel Corporation)