by Anthony Toprac, W. Jarrett Campbell, Yield Dynamics Inc., Austin, Texas
Statistical process control has played a large role in semiconductor manufacturing, but sophisticated physically based models recently have revealed an underlying structure rather than random processes with unknown causes of variation. Now, advanced process control is emerging as the next level of controlling wafer fab effectiveness and a manufacturer's competitiveness. Indeed, implementation of these new software control systems in the fab may be one of the most critical activities that manufacturing engineers can engage in today.
Advanced process control (APC) is gaining widespread acceptance as an important tool in high-volume semiconductor manufacturing. This trend is evidenced by the sharply growing attendance of the annual Sematech AEC/APC Conference (Fig. 1) .
Figure 1. Annual attendance at the Sematech AEC/APC Conference (*attendance projected for 2001).
This sudden upsurge of interest in APC systems can be attributed to several well-known trends in semiconductor manufacturing, including high capital costs, accelerating technology requirements, and lagging tool capabilities. The paradigm of statistical process control (SPC) has become insufficient for control needs generated by these trends. For instance, SPC out-of-control situations usually result in tool downtime. The tool remains out of operation until maintenance is completed and the tool is re-qualified for production. At the current costs of semiconductor process tools, this downtime carries a substantial penalty in lost productivity. Worse still, the frequency of SPC out-of-control events increases sharply when aggressive process requirements exceed tool capabilities.
One of the most cost-effective methods of addressing these issues is the use of APC technology. APC refers to a wide range of technical areas, from factory-level software systems to specific sensors implemented in particular tools. But the bottom line is that automated control systems are a multiplier of fab productivity available at a small fraction of today's wafer fab capital costs.
Real-time vs. batch control systems
An important distinction in control systems is real-time vs. batch control. Real-time systems conceptually provide instantaneous control action, although in practice real-time controllers execute on a synchronous time interval governed by system clock speed. Real-time control systems require instantaneous feed-forward and feedback information from real-time measurement systems. In semiconductor manufacturing, real-time control is part of production tool architecture, providing accurate and repeatable recipe execution.
Analysis and modeling for real-time controllers are usually performed using frequency domain analysis to determine tuning constants for the proportional-integral-differential (PID) control algorithm. This methodology dates back to the 1930s  and has been the workhorse of real-time control applications in the chemical process industry. More recent developments in chemical process control involve model predictive control (MPC), a method used for plant-wide control of refineries and chemical production units. Typical MPC algorithms employ linear multivariable models of the plant to optimize a control object over an infinite horizon of time steps .
MPC algorithms are finding their place in semiconductor manufacturing tool real-time control systems. A prime example of this is batch furnace temperature control. Any manufacturing tool that requires dynamic control of multiple, coupled process variables is a prime candidate for an MPC algorithm, and this methodology will become increasingly commonplace in internal control systems of wafer fab tools.
In contrast, batch control system execution is clocked on asynchronously generated events, such as batch process initiation or post-process metrology completion. In semiconductor manufacturing, these events occur at widely varying time intervals from seconds to hours. Design and implementation of batch control systems is handled by manufacturers. This ranges from manual adjustments based on SPC rules to more advanced, automated techniques such as run-to-run control. Run-to-run control describes batch feed-forward feedback automated control using pre- and post-process metrology to change process recipes.
The concept of automated run-to-run control has only recently been accepted in semiconductor manufacturing. Ironically, a major inhibitor to automated run-to-run control was the control technique that provided major manufacturing control advances in the 1980s. This technique is SPC. It characterizes measured process results as random variation about a mean value that, under normal operation, represents the process target. Since results away from the process target, or the error, are representative of random process noise, any attempts to apply feedback corrections based on the error only amplify random noise, worsening process control. This concept was circulated by Deming  as part of his promotion of SPC methodology, and is frequently demonstrated in courses teaching SPC methods by using examples of statistically random processes.
Although this conclusion is mathematically correct, feedback of corrections based on process error has been shown to improve process control significantly in many wafer fabs [5-8]. Figure 2, which depicts a typical signature for process data from a semiconductor manufacturing line, illustrates this concept.
These data, rather than being randomly distributed over time, actually consist of a deterministic low-frequency component, indicated by the dotted line, superimposed on a random noise signal. This low-frequency component is correctable by employing feed-forward and feedback information in an automated run-to-run control scheme. Run-to-run controllers actively provide lot-by-lot recipe adjustments based on a deviation from target. Elimination of the low-frequency process error by run-to-run control is shown in Fig. 3. The output in Fig. 3 represents a normally distributed random process that can be handled well by SPC alarming techniques for nonstandard process behavior.
In contrast, SPC-based feedback control typically employs simple logical rules to determine whether the error represents a true bias as opposed to random variation. The variance of the data in Fig. 2, however, includes a deterministic low-frequency variation. A SPC-based controller attributes this deterministic signal to random variation and does not provide any correction. Moreover, SPC methods of feedback control are not typically automated and require manual intervention. Often, the result is persistent off-target operation. This is in contrast to processes under run-to-run control, which automatically centers the process on-target. In addition, the degree of manual control with SPC methodology is often based on heuristic rules and human judgment, whereas automated run-to-run control is calculated by mathematical models correlating control inputs to process outputs.
Since semiconductor manufacturing is based on batch processing, batch control techniques such as run-to-run control have tremendous potential for improving wafer fab operations. Future fabs will have highly integrated batch control systems, using feed-forward and feedback information across multiple, distributed operations. While lot-to-lot control is the current technology, wafer-to-wafer control will become implemented in 300mm fabs within a few years. This requirement will be driven by high wafer and manufacturing tool costs, and will be made possible by the integration of post-process metrology in process tools.
While today the implementation of run-to-run controls lies in a manufacturer's domain, the advent of wafer-to-wafer control will require batch control methods to be implemented by tool suppliers as part of their product offerings. Tool-based wafer-to-wafer control, however, will still require integration with the manufacturer's fab-wide batch control system. This integration will supply the tool-level batch controller with wafer-specific feed-forward information from previous processes as well as updates to process targets based on feedback from downstream product measurements.
Fault detection systems
Fault detection and classification (FDC) systems are another major component of APC software solutions. Fault detection is the detection of abnormal process operation. Classification provides an assignable cause to this excursion. FDC software systems provide an automated means of detecting, alarming, and assigning cause to abnormal process operation.
The detection of abnormal operation is the most important benefit that a SPC system provides. SPC systems, however, are geared to monitoring single signals. The result has been an explosion in the number of SPC charts needed to monitor the high volume of operating data in today's wafer fab. Each of these charts requires maintenance and monitoring to be effective.
A primary feature of automated FDC software systems is the ability to monitor processes on a multivariate basis. Rather than independently monitoring a large number of process variables, a single process statistic is monitored. This single statistic results from an FDC model with multiple process variables as input and can be monitored using a single SPC chart. In addition, a multivariate model improves the sensitivity of excursion detection by including interactions between process variables. For example, a change in the correlation between two process variables can be detected by multivariate FDC while SPC monitoring shows these variables individually to be within statistical control.
Sophisticated multivariate FDC systems have tremendous potential in monitoring tool performance and diagnosing tool faults. The large number of variables that must be monitored in a process tool requires tool suppliers to embed multivariate FDC systems in their control software architecture. In addition to improving tool operation, FDC systems can operate at the process level with signals from multiple process operations combined to monitor process health. Multivariate modeling of process, metrology and electrical test data will provide future fabs with an on-line quality monitor of everything from individual lots to the entire manufacturing line.
Software system architecture, integration
Over the last five years, several suppliers have emerged with commercial software for APC applications. Much of the focus on APC software is on the development and implementation of APC framework systems. A framework is software infrastructure that facilitates efficient implementation of control applications. There are two major requirements for this framework.
First, the framework should embody all the functionality common in wafer fab control applications. For instance, data storage is a functionality required by all run-to-run control applications, so a database component is a requirement for an APC framework.
Second, this framework should integrate with other software components in the fab, including the manufacturing execution system (MES). Integration is a major component of implementing a new control system. Providing an infrastructure that facilitates integration is a major benefit of a framework.
Seamless integration of software components requires well-defined standards for component behavior and communication protocol. The first attempt at establishing APC framework standards was undertaken by the APC Framework Initiative (APCFI), a joint development project between Advanced Micro Devices, Honeywell, and Sematech performed under a supporting grant from the National Institute of Standards and Technology (NIST) Advanced Technology Program.
Begun in 1996, APCFI resulted in a publicly available design specification with standards for software component interactions . Framework functionality was divided into multiple components with the idea that behavior and communication standards would foster plug-and-play capability among multiple suppliers. These standards would allow the manufacturer to choose the best of breed for individual components to assemble a complete APC framework.
The process of software standard acceptance, however, greatly reduced the granularity defined in the NIST APCFI project. Negotiations among interested parties resulted in the Semi E93-0200 Provisional Specification for CIM Framework Advanced Process Control Component .
The Semi E93 provisional specification is limited to describing just two software components a "control executor" and "algorithm executor." Despite the lack of detailed functionality in this specification, it does provide a delineation of a standard interface between the MES and the APC framework. Actual use of this standard interface, however, requires that the MES comply with the Semi CIM standards , which recently have been incorporated in commercial MES products.
Although the E93 APC Framework specification provides in theory a means of seamless integration with a Semi CIM standard MES, it does not address issues involving MES integration to a tool. While these integration issues are much broader than APC functionality, the difficulty of MES-to-tool integration often causes great delays in implementing APC applications.
The major obstacle of MES-to-tool integration is the lack of robust SECS-GEM standards  for tool communication. Lack of detail in this standard has led to great variation among implementations of the SECS-GEM protocol. The result is that implementing MES-to-tool communication requires protracted cycles of characterization, custom coding, and testing for every tool delivered to the fab. In addition, specific SECS-GEM functionality required by APC application may not have been implemented in a tool's software. A prime example is the lack of remote parameter setting ability, a requirement for APC to change process recipes on a run-to-run basis; and even the most carefully crafted MES software cannot avoid problems generated by tool software lock-up due to weak implementations of the SECS/GEM communication port.
Clearly, the current status of APC framework software and software integration leaves much room for improvement. Legacy MES systems require customized integration to an APC framework, as well as custom implementations to communicate with tools. Even state-of-the-art MES and process tools are a long way from the plug-and-play environment envisioned in the Semi CIM framework specification.
Nonetheless, evolution in software standards will eventually bring plug-and-play interoperability. This paves the way for horizontal integration of APC applications into a fab-wide control system and vertical integration of APC functionality into a manufacturer's business enterprise systems. Future wafer fabs will have a similar appearance to today's chemical or power plants, with a small operating staff monitoring the production line from a central control room while a handful of technicians monitor and service a fully automated manufacturing floor.
Suppliers currently offering APC software are distributed across a variety of business entities. A large percentage of APC software solutions represent point solutions built and supported by a manufacturer's in-house IT and process module resources. Similar point solutions are offered commercially by suppliers that specialize in off-line analysis software for particular process areas such as photolithography. In addition, hardware suppliers are beginning to offer APC software as value added control solutions with a purchase of their tool sets.
Suppliers that base business solely on software solutions for semiconductor manufacturing, while few in number, are best positioned to meet the needs of a wafer-fab customer. Manufacturers have always strongly relied on their ability to select best-of-breed production tools for their new fabs. Today's advanced wafer fabs have a similar requirement for plant software systems. Linking plant-wide software solutions such as APC to particular hardware platforms locks a manufacturer into a selection that can be second best on either the hardware or the software end. In addition, software suppliers that base all of their business on control solutions have the most incentive to push control technology development and, consequently, are staffed with the most experienced and highly qualified control engineers.
The future of semiconductor manufacturing's APC software can be determined by considering advanced control software in the Chemical Process Industry (CPI). The CPI has actively used APC software for more than two decades. Advanced control solutions in the 1970s were provided by in-house experts who developed, tested, and implemented APC software in chemical plants and refineries. Throughout the 1980s, multiple companies provided these software solutions commercially, while the manufacturer's in-house capabilities simultaneously declined due to downsizing and attrition. Finally, the 1990s saw consolidation of these companies into three large, well-established companies that account for virtually all of the commercial CPI APC software business.
The APC software business for semiconductor manufacturing, lagging by 20 years, is on a similar track as the CPI industry. Currently, the bulk of APC software solutions operating in semiconductor fabs are the result of in-house development. Within the last three years, however, multiple companies have entered the market with APC software products. We can anticipate that over the next decade, most of the APC software business will move from in-house to commercial providers. These commercial suppliers will, in turn, consolidate to form a few large, well-established software houses that service the process control requirements of semiconductor manufacturers.
What the future holds
Semiconductor manufacturing presently lags the chemical process industry in advanced control technology. Ironically, the CPI's lead was made possible by the evolution of technology in semiconductor manufacturing. The SPC concept of manufacturing as a collection of statistically random processes has played a large role. In recent years, sophisticated physically based models have revealed an underlying structure that belies the treatment of semiconductor manufacturing processes as random processes with unknown causes of variation .
In any case, the return on investment described by publications of APC applications  indicates that manufacturing competitiveness will be determined by the effectiveness of control systems.
1. B. Van Eck, Sematech, personal communication 4/2/01.
2. Seborg, et al., Process Dynamics and Control, John Wiley & Sons, 1989.
3. K. Muske, J. Rawlings, "Model predictive control with linear models," AIChE Journal, Feb. 1993.
4. W.E. Deming, The New Economics For Industry, Government, Education, M.I.T. Center for Advanced Engineering Study, 1993.
5. S. Bulter, J. Stefani, "Application of Predictor Corrector Control to Polysilicon Gate Etching," Proceedings of the American Control Conference, 1993.
6. W.J. Campbell, "Model Predictive Run-to-Run Control of Chemical Mechanical Planarization," University of Texas at Austin, 1999.
7. C.A. Bode, "Run-to-Run Linear Model Predictive Control of Lithography Overlay," AEC/APC Europe Symposium, 2000.
8. J.C. Stuber, "The Effect Of Model Form and Controller Gain on Run-to-Run MBPC of LPCVD Films in Horizontal Furnaces," AEC/APC Symposium, 1999.
9. Advanced Process Control Framework Initiative (APCFI) Project: Detailed System Description, International Sematech Tech, Transfer No. 99053736A-TR, June 30, 1999.
10. Provisional Specification for CIM Framework Advanced Process Control Component, Semi E93-0200, Sept. 1999, www.semi.org.
11. Provisional Specification for CIM Framework Domain Architecture, Semi E81-0600, June 1999, www.semi.org.
12. Provisional Specification for Semi Equipment Communications Standard 1 Message Transfer, Semi E4-0699, June 1999, www.semi.org.
13. Ziger, et al., "Generalized Characteristic Model for Lithography, Application to Negatively Chemically Amplified Resists," SPIE Microlithography Proceeding, 1991.
14. J. Baliga, "Advanced Process Control: Soon to be a Must," Semiconductor International, July 1999.
Anthony Toprac received his PhD in chemical engineering from the University of Texas at Austin and is a registered professional engineer. He is VP of APC Solutions at Yield Dynamics Inc. and the director of Yield Dynamics' APC Development Center, Austin, TX 78731; ph 512/257-9500, fax 512/257-9503, e-mail email@example.com.
W. Jarrett Campbell is a graduate of the Georgia Institute of Technology and received his PhD from the University of Texas at Austin. Campbell is a member of the technical staff of APC Solutions at Yield Dynamics Inc.