APC: 2004 Technology and Market Update

Much continues to be written about the technology and benefits of advanced process control (APC), and as the industry finally emerges from the downturn, APC has become a requirement for every new fab or capacity expansion around the world. In such an environment, you might think that the specification, evaluation, selection and deployment of an APC solution and its supplier(s) would be a straightforward matter — but you'd be very wrong. In fact, as APC has been more broadly deployed in leading-edge wafer fabs, the situation has become much more complex. This is partly due to the inherent connectedness of these applications with other systems in the fab, partly because APC technology itself has evolved, but also because of the changes wrought in the commercial market over the past two years.

Accordingly, this article highlights some of the major technology and market factors that you should consider when approaching this domain.

One sign of a maturing market segment is consistency of terminology and definitions, and in this regard, the semiconductor APC market has come a long way. Variations of the technology have been applied across a wide range of processes, time spans and product material groupings — and each situation has its special set of challenges and solutions. This is one of those cases where a picture is worth at least a thousand words, so please see the Figure for a summary of how the different APC sub-segments relate.

Variations of APC technology have been applied across a wide range of processes, time spans and product material groupings — and each situation has its special set of challenges and solutions.

All of the top 20 semiconductor manufacturers now have some form of production APC capability in place for the critical process steps, and support active R&D programs to guide future deployments. The APC technology leaders among them have implemented fab-wide infrastructures and distributed application frameworks to support multi-process, "cascaded" and hierarchical control applications that use information from many disparate sources and timeframes. Some have integrated their engineering analysis, yield management, production data mining, and other offline systems to diagnose control problems and target their APC investments. Some have included final electrical test and other downstream information in long-cycle control loops to adjust some of the earliest processing steps. Others have factored the need for fresh production data to support the controllers of low-running devices into their dispatching algorithms. Still others have implemented adaptive sampling algorithms that vary the trace/metrology data collection plans and even the production flow based on controller results. And others have automated systems that monitor the behavior of the APC applications themselves. And the list goes on...

The message here is that APC is fundamentally a continuous improvement discipline that impacts a wide variety of manufacturing technologies and job functions. When viewed in this light, a comprehensive, well-planned APC strategy can be one of the most promising and pervasive initiatives a company can take.

The majority of production run-to-run (R2R) control algorithms are based on linear models that relate pre- and/or post-process product metrology values (film thickness, critical dimension, alignment error vectors and the like) to the process recipe parameters (process time, exposure dose, RF power, stage correctables, etc.) that have the most direct impact on the desired result. Applying the proper model for a given situation (product, layer, tool/chamber, reticle set) is accomplished by using these "context" values to select a specific set of coefficients from a potentially large database of active models. Moreover, since the fidelity of this type of control model depends on having enough recent history for each context combination, it can be problematic in high product mix environments; in fabs that must support a great deal of process development or that experience frequent tool maintenance or calibration; and other manufacturing situations that are relatively dynamic for whatever reason.

Likewise, the algorithms in most of today's fault detection/classification (FDC) systems use univariate or multivariate statistical techniques and experimental trace/metrology data sets to characterize tool/process behavior and develop a set of fault models. These fault models are then applied to production trace data to determine whether a tool has drifted out of a valid process window and may be creating scrap (or worse). However, since these models have no direct mathematical relationship to the actual physics, chemistry and mechanics of the processes being monitored, they are very sensitive to the operating conditions and production context in which they were developed, and as process windows narrow, the level of false positives grows. Relax the fault criteria too much, and real problems may be missed. If this situation is not corrected, the production users will soon lose faith in the system's ability to accurately diagnose problems, and perhaps cease using it altogether.

Finally, when R2R and FDC solutions are both deployed in a given process area, they must be able to interact and update/share tool state information, or they will be working at cross purposes.

The solutions to these problems will take many forms, including multi-input multi-output (MIMO) control models that consider more product/process parameters simultaneously to select optimum operating points; so-called "first principles" or "physics-based" techniques that include mathematical equations of the process conditions and resultant material transformations, and which would therefore be much less dependent on production context information; process engineering tools that allow relevant domain knowledge to be captured, structured and applied; and combinations of all the above.

The global community of SEMI Standards volunteers and support staff have been busy during the recent downturn, creating a comprehensive and complex family of standards for accessing and using potentially massive amounts of detailed tool/process data. It's a bewildering array of technologies and documents: common equipment models, XML usage, data acquisition management, Interfaces A/B/C, equipment self-description, authentication and security, data quality, process control system interfaces, modules, equipment performance tracking and quality information parameters, integrated metrology module interfaces and tool micro-cycles (E134, E120, E125, E133, E126, E116, E132, E127, Doc. #s 3571, 3905, 3985, ...).

The theme of this initiative, however, is relatively simple: The collection, analysis and visualization of detailed tool and sensor data leads to a deeper understanding of tool/process behavior, which in turn leads to improved equipment and more effective manufacturing methods. From an APC perspective, this information will support the evolution of more robust algorithms that are needed to overcome the limitations of the heuristic and statistically based techniques prevalent today. Moreover, the unprecedented availability of this kind of information will create additional demand for software tools and productivity improvement applications that we can barely imagine at this point.

The limiting factor will be how quickly and broadly these standards can be adopted across the industry. So another important thing to remember is that, unless you're in the packaged standard interface software business, "you shouldn't try this at home" — spend your effort figuring out how to best use this data, not how to build the plumbing.

Fab-wide APC systems are now commonplace — but the landscape has changed. The industry has moved beyond the original goal of a plug-and-play environment in which modular control applications could snap into the well-defined sockets of a single, fab-wide, industry-standard APC infrastructure. The length and severity of the downturn and the major investments made by end users in their existing systems have created a legacy APC system situation that will be with us from this point forward. Moreover, much of this work has been done on private (read "home-grown") or semi-private systems, so very little of this investment has trickled down into the supplier community. In such a setting, a "peaceful co-existence" product architecture strategy is the most viable approach for APC suppliers hoping to penetrate major end user accounts.

This means that major analysis and control applications must be deliverable as standalone systems with well-defined integration points and robust infrastructures that provide high availability at production volumes. Moreover, these applications must provide the flexibility and IP protection required by the do-it-yourselfers, which make up an increasingly large portion of the APC market. This is a pretty tall order for any software development team, but especially the small to medium-sized companies that have traditionally served this industry.

Another approach, of course, is to specialize in a particular set of process technologies, making sure your algorithms can be implemented and packaged to leverage someone else's infrastructure. But how do you go about deciding whose to pick?

The silver lining in all of this is that no dominant supplier has captured much of this market, so there is still plenty of headroom for companies that have the insight, resources, patience and will to serve this important segment. As a result, the semiconductor APC market is still attracting investment from innovative suppliers — from other industries as well as our own.

To support the continually changing requirements of a modern wafer fab, no fixed set of out-of-the-box features that match the list on someone's RFP would be sufficient. Even if a complete solution was available at some point in time, no standard product can hope to keep up with the pace of innovation that will be set when this industry's production and process engineers are armed with more and more data. Consequently, APC solutions must also have the inherent flexibility necessary to enable their providers (internal and/or external) to respond quickly with new control application capabilities — the combination of these two attributes is often called agility. The architecture must be considered at the very beginning of a system design.

There are many ways to measure this capability throughout the life cycle of an APC system, from validation pilot to full fab rollout, but all of them factor in the elements of change and time. "Better never than late" has been used to emphasize the importance of time-to-market in the highly competitive IC market; these same pressures apply to a fab's APC development, deployment and support teams, so agility should be an explicit evaluation criterion for any future APC system.

APC is now a mainstream chipmaking technology, and considered essential for the 130 nm, 300 mm nodes and beyond. The options are more numerous, the benefits more significant, and the decisions more complex than ever before. But the most costly decision is to do nothing — so gather your requirements, get help where you need it, embrace your suppliers as partners in this venture, and take control.