Advanced Process Control: Soon to Be a Must

John Baliga, Associate Editor -- Semiconductor International, 7/1/1999

Advanced process control (APC) has been used in semiconductor manufacturing to a small degree over the past decade. At present, it is most commonly used in chemical mechanical planarization (CMP) processes (Fig. 1). APC holds a great deal of potential for reducing scrap, increasing time between preventative maintenance (PM) operations and making more devices bin out at higher performance levels by tightening the process distribution. That potential has gone largely untapped, and tapping into it soon will be a necessity.

APC falls into two main categories: run-to-run (R2R) control, and fault detection and classification (FDC). Run-to-run control of a process on a tool uses data from outgoing and incoming wafers, in combination with a model of the process in that tool, to adjust process parameters. Fault detection is the prediction of an imminent tool failure, and fault classification determines the cause of that failure. Performing the necessary maintenance without sacrificing wafers is the goal of fault detection and classification.

The 1997 National Technology Roadmap for Semiconductors (NTRS)¹ points out the need for advanced process control, though it uses different terminology. In the section titled Sensor-based Metrology for Integrated Manufacturing, it states, 'Given the continually increasing process requirements for future devices and structures, coupled with the equipment reliability, cost and uptime requirements for 200-450 mm wafers, semiconductor manufacturing is slowly moving from fixed process recipe, open-loop control to closed-loop control via sensor-driven model-based integrated manufacturing (SDMBIM).'

The NTRS goes on to describe how metrology will have to change: 'One key pacing item is the transition away from off-line metrology that has been the mainstay of semiconductor manufacturing. Currently, metrology is shifting toward in situ sensors, as these can provide in-time data for active process control and the elimination of wafer misprocessing. For 200 mm (and even more for the 300 and 450 mm wafers), misprocessing is becoming prohibitively expensive.'

It may be difficult to use situ sensors for full real-time control of all processing. The concept is used for endpoint detection in some process steps, but it has more potential. Some have expressed a desire to detect the start of a misprocessing event in real time to make corrections immediately, saving the wafer being processed. This is a very ambitious goal that may need to be met when 450 mm wafer processing is developed, or possibly sooner, depending on the process step. For now, using run-to-run control to keep wafer states closer to their nominal parameters, and fault detection to predict and avoid misprocessing, can make more ICs achieve peak performance and save many 200 and 300 mm wafers from becoming scrap.

Not just more SPC

APC is not just another form of statistical process control (SPC), nor a replacement for it. SPC is used to control a process in the face of random variations. APC is used to control processes in the face of systematic variations. For control of random variations, active feedback control would be a bad choice. APC provides active control of systematic variations, and it has produced very promising results.

Some variations that may seem random are really systematic. Other variations are known to be systematic, but they are treated as random for control purposes. Many process tools drift over time, which, by definition, defies the use of statistics that assume a constant mean on outgoing wafer-state characteristics. The feedback methods of run-to-run control can be used to undo that drift with recipe adjustments; then SPC methods then could legitimately be used on the outgoing wafer-state characteristics. Feedforward methods would help make adjustments at downstream tools to fix characteristics that still may be in spec, but not optimal.

Run-to-run control

Run-to-run control uses metrology data taken at each process step to adjust process recipes on a run-to-run basis. A run may be a batch of wafer lots, a single lot or even a single wafer, depending on the particular needs of the process step and the fab.
Run-to-run control can be done with stand-alone tools, although using in-line sensors measuring one or more wafer parameters would be faster. Run-to-run control keeps wafer-state parameters close to their nominal values using both feedforward and feed back techniques. Data taken immediately after a process step on a particular tool can be fed back to adjust the recipe for the following run. This requires adjustable recipes and a reasonably accurate model of the process as performed by that tool. With feedforward, data are sent to the next tool to adjust its recipe. This can fix a problem before it gets bigger.

Fig. 1 CMP processes are among the first to aggressively use run-to-run control with integrated sensors. (Source: Applied Materials)

The industry has lived with drift and variation from its process tools for years, but the time is coming soon when these variations can no longer be tolerated (see The Barefoot Shoemaker). Commenting on some of the work done on the 1999 International Technology Roadmap for Semiconductors, Daren Dance of Wright, Williams & Kelly (Austin, Texas) said, 'The one thing that is very, very imperative is the tolerance for error in processes is evaporating. To maintain cost-effective manufacturing, we've got to have control schemes that will allow us to hit nominal consistently with no drift.' He also noted that since upcoming semiconductor products have such high value, 'a little bit of scrap elimination can justify an awful lot of improvement in process control, if and only if it eliminates that variance.'

Figure 2 shows how both feed forward and feedback can improve a CMP process.² The effects depicted can apply to every step in the overall process. In some processes, endpoint detection may make feedforward unnecessary, but the feedforward method may be easier to implement.

In many cases, the algorithms for implementing the control are relatively simple, and the tool or process modeling does not need to be very sophisticated. The critical path is in taking the data and accessing it. Scott Hawker of Motorola said, 'A lot of our process characterization work is to squeeze as much data as we can through that 9600 baud SECS port, which is a problem, obviously, and use that to capture a database for analysis. We get our process characterization from the outside of the tool.' To that, Karen McBrayer of Motorola added, 'If you look at a lot of the equations that have been published in the literature to date, they're really simple things. Most of them are one line. Only a few of them have more complication than that.'

The other side of that coin is that run-to-run control should only use simple models Commenting on run-to-run control for overlays in lithography, Patrick Lord of KLA-Tencor said, 'Depending on whether you have a stepper or a scanner, there could be at least six to eight alignment parameters that you need to adjust. Fortunately, for the first order correctable models, the parameters are orthogonal, so you can feel fairly comfortable adjusting any or all of them.' When it comes to things like tilt, it is not so simple. 'For tilt you would design an experiment with a test wafer, figure out the trapezoidal errors, determine the tilt and compensate for it,' Lord said. 'Trapezoidal effects, being second order in nature, can definitely couple with the terms in your first order model. In my opinion, tilt compensation would not be something for run-to-run control. It might be a part of a PM, monitored with SPC.'

Fig. 2 Though either the feedback or feedforward part of run-to-run (R2R) control can be a great help, both are required for effective control. For a CMP example without endpoint detection, thickness measurements are taken before and after polishing (a). When only feedforward is used, variations due to the incoming thickness are corrected, but the changing removal rate may leave the outgoing thickness out of spec (b). When only feedback is used, the polish time is changed to compensate for the changes in the removal rate, but the variation of the incoming thickness is preserved (c). With both feedforward and feedback, the ouptut thickness stays near its nominal value (d), and the SPC can be used to refine the control. Since R2R is model-based, it is possible to automatically compensate for events such as a pad change (e). (source: SEMATECH)

Sensors

Metrology must be performed at every process step to make run-to-run control possible, and data from tool subsystems must be available to make fault detection and classification possible. Taking wafers to a stand-alone metrology station may be too time-consuming to perform measurements on every wafer for run-to-run control. Run-to-run now is done with stand-alone tools on a lot-to-lot level.

Measuring critical dimensions on gates may require taking wafers to stand-alone tools, but other parameters can be done with simpler equipment. CMP processes were the first to use add-on sensors to perform wafer and thin film measurements for run-to-run control. Nova Measuring Instruments' (Rehovoth, Israel) NovaScan 420 was created to make those measurements on wet wafers. Nanometrics (Sunnyvale, Calif.) recently offered its NanoSpec 9000 system for integrated film thickness measurements for a number of applications including CMP. The Ellipson sensor from NanoPhotonics AG (Mainz, Germany) is a compact ellipsometer available in an OEM version. On-Line Technologies (East Hartford, Conn.) makes Fourier transform infrared (FTIR) equipment for various fab applications, including thin film thickness.
One example of integrated measurement other than CMP is the control of a silicon epitaxy process with one of On-Line's FTIR sensors on the cooldown chamber of an Applied Materials cluster tool (Fig. 3). 'In-line metrology has a number of benefits for the epi process,' said Grant Imper of Applied Materials Epi Division. 'It can reduce monitor wafer costs, which will be particularly significant at 300 mm. Customers have also seen benefits in being able to more quickly tune in new process specifications. In addition, when integrated to the software system cell controller, in-line metrology provides an extensive database and realtime system controllability for improved wafer-to-wafer yields.

Created last year, the Integrated Measurement Association (IMA, www.integratedmeasurement.com) promotes the use of APC in semiconductor manufacturing by helping create cooperative solutions among technology providers. As the name indicates, its work involves the integration of sensors. The IMA plans to foster better communication among end users, suppliers of sensors and software, and fabrication tool manufacturers to identify important problems and facilitate cooperation in their solutions. More information about the IMA, its member companies and its efforts is available on its website.

Tool subsystem sensors also are changing to reliably provide more data. For example, a mass flow controller (MFC) controls a specific gas flow. This is not a part of process control, but the MFC can provide information essential for fault detection. Many MFCs have their own processors and digital communication capability using systems like DeviceNet, LonWorks and Smart Distributed System.

This digital signalling capability is useful for several reasons. First, it is readily compatible with data acquisition and analysis software packages. Digital communication also is more reliable in a fab setting. Joe Maher of MKS Instruments (Andover, Mass.) said, 'For typical mass flow controllers, we're looking at millivolt type of levels to set to get a particular flow, for instance. Most process tools, if you look at PVD, plasma, CVD or any other kind of tool, are dealing with pretty hefty electrical loads. They may have a couple thousand watts of rf or a couple thousand watts of heaters turning on and off, and there are always ground loops, even though you're not supposed to have them. It doesn't take much to knock that millivolt level signal off a bit, and therefore have a wrong set point going to the mass flow controller and consequently a wrong flow. The nice part about digital technology is that instead of looking at a millivolt level signal, you're looking at a 24 volt digital signal in the case of DeviceNet. Corrupting that signal is pretty difficult.'

Maher also noted that with digital technology, an MFC can continually report more data values, such as the valve voltage, which indicates the degree to which the valve is open. The MFC uses this data in controlling the flow. If the valve has to stay farther and farther open to achieve a constant flow, that may indicate the valve is getting clogged. The tool and factory level controls certainly do not get to change the valve position directly, but tracking data like that can alert fab personnel to take the tool off-line and perform maintenance before it crashes, saving a lot of money in scrap wafers.

Data communication and analysis

The slowest link in migrating to APC is upgrading equipment communication capability. Though the SECS-I and SECS-II communication standards have been in place for quite some time, there are still many tools in many facilities that do not have this capability. Most equipment with this capability uses the more familiar RS-232 standard instead of the much faster HSMS ethernet standard. For the data volumes required for APC, many say the HSMS data buses will be minimum requirements, as well as sensors and subsystems that can keep that HSMS cable busy.

Equipment communication standards are not the limiting factor. Jack Ghiselli of GW Associates said, 'I think many pieces of equipment don't work well with APC. It's not so much a SEMI standards lack, as a lack of good implementation of existing standards.' He noted that when CMP equipment started using APC, the recipe adjustment capability was in place but not hooked up with the tool communication system. 'When people first started to do APC applications on CMP tools, they found that the SECS interface did not provide the ability to tweak the recipe. The builder of the equipment had just never thought about that issue,' Ghiselli said.

Fig. 3 More sensors are being designed for in-line run-to-run metrology. This FTIR sensor is mounted on the cooldown chamber of a silicon eppitaxial reactor. (Source: Wacker Siltronic Corp.)

Carl Fiorletta of Adventa Control Technologies said that in many places, even the minimum communication capability is lacking. 'At the lowest level, we still see GEM/SECS interfaces on process machines that don't function to the SEMI standard, or don't function in accordance with the specifications provided by the equipment manufacturer. Between the tool and the CIM System, we typically don't find a robust communications network between tools and the factory LAN to support data collection from the tools and recipe download to the tools. When we install APC in a production factory, the last thing we do is install APC. First we must create the proper environment for APC to function,' said Fiorletta.

Installing the necessary communication capability will require some effort and expense. One thing that can make this easier is the APC Framework, now a part of the CIM Framework, developed by AMD, Honeywell, SEMATECH, Object-Space and Oak Leaf Engineering under a program partially funded by NIST. The APC Framework is designed to make the integration of APC software applications standard, so the choice of sensors and control algorithms are made according to a fab's particular needs and goals rather than compatibility issues. The Framework is platform independent, making it possible to transfer control algorithms and models more easily (see The APC Framework and Process Technology Transfer).

Fault detection and classification

Advanced equipment control fits into two main categories: real-time control and fault detection. Real-time control is mostly endpoint detection. Fault detection can occur at many levels within a process tool.

Many individual components in a tool have their setpoints, and alarms go off when they run outside them. These parts can be changed out, but usually by the time that is done, some wafers have been sacrificed. Model-based control allows the use of a predictive alarm, alerting fab personnel to take the tool down before any wafers are misprocessed. This requires a reasonably sophisticated tool model. A variety of software tools are available for tracking data and performing model-based control (Fig. 4).

Brad Van Eck of SEMATECH said fab personnel are doing fault detection and classification manually, after the fact. 'They have some kind of a data acquisition system hooked up to the tool, and they just suck the tool dry of data and stare at it. Eventually they say, 'Look, this data predicted when that maintenance should have occurred. I have this data from five different PMs on the same tool. This time we should have started the PM a little earlier; this time we should have let it go for a while.' They are staring at the data and trying to figure out if the faults that they actually observed were evident in the data and that they could have had some early warning that it was coming.'

Fig. 4 A variety of software tools exist for fault detection and classification. (Source: Triant Technologies)

This would indicate that many facilities have the necessary knowledge to implement an automated fault detection and classification system. Van Eck noted that some have started doing so -- integrating sensors and using data from them as well.

When a suitably sophisticated model of the tool is not available, multivariate SPC can be used. Quite often, univariate SPC charts are kept for multiple machine parameters, and as long as each parameter stays within its 3-sigma limits, the tool is thought to be under control. Unfortunately, there are plenty of cases in which each univariate parameter is within its control limits, but the joint probability distribution is not, even if all variables are independent of each other.

There are a variety of ways to implement multivariate SPC fault detection. IBM Microelectronics (Burlington, Vt.) has used a straightforward Hotelling T-squared statistic on the variables.³ All methods involve modeling the tool, and most models are developed from data taken during operation. This approach has another benefit: a model developed for a good tool can be used to bring up a troublesome tool more quickly. Ray Bunkofske of IBM Microelectronics gave this example: 'Let's say you've got 25 of tool X. Probably three of these tools run very well: they provide good results and they work every time. You'll have other tools that are not quite so good, and a couple of poorly performing tools that you can't keep up more than a day at a time. Build a model from your good tools, and put that model on all the other tools. They're going to trip out every third lot or so, but you'll find out that a certain MFC isn't quite good, or that the pressure controller isn't quite good or that the PID constants on the rf generator were wrong, et cetera. Start tuning them up and pretty soon all your tools look like your best tool.'

Classifying the fault is just as important as detecting it. First, if a tool is to be brought off-line for maintenance on short notice, knowing what to fix can help bring it back on-line sooner. Classification also helps identify the responsible party. Depending on the fault, correction could be the responsibility of the equipment or process engineer.

This delegation of responsibility may be one of the more difficult barriers to implementing APC. If the equipment supplier provides the APC capability, the supplier would be responsible for maintaining that capability, depending on the arrangement with the end user. If the end user implements it, the maintenance and warranty arrangements are more complicated. If a third party is brought in to integrate that system, they can become even more complicated. The benefits of an APC system can make this extra trouble worthwhile.

Conclusion

Equipment suppliers already provide real-time active control of equipment state parameters. Soon, real-time active control will be a necessity for cost-effective semiconductor manufacturing. There is a great deal to be gained from advanced process control in terms of reduced rework, more devices working near their theoretical peak performance and reduced scrap. Any one of these advantages can pay for the implementation of APC within a few months and go on to make the fab that uses it more profitable. Table 1 is a brief list of some of the published studies showing specific advantages to using APC.

There seems to be enough knowledge about processes to implement APC widely, but data collection and communication need to be upgraded before that can happen. There is a trend for tool subsystem controls to be adaptive rather than hard wired, and that may drive data communications to improve from one side while the need for APC drives it from the other.

With technology improving even faster, integrating tools and controlling processes need to become less of a chore so that fab engineers can concentrate on developing new processes and matching their production to ever-changing demand.

References

1. National Technology Roadmap for Semiconductors: Technology Needs, 1997 Edition, Semiconductor Industry Association.
2. Advanced Process Control Framework Initiative v 2.0 Sematech Technology Transfer #97063300B-ENG.
3. R. Bunkofske, 'How to Use Real-Time Process Monitoring to Improve Yield and Reduce Manufacturing Costs,' SEMATECH AEC/APC Symposium X Tutorial, October 11-16, 1998.
4. D. Gerold, R. Hershey, K. McBrayer, J. Sturtevant, 'Run-to-Run Control Benefits to Photolithography,' SEMATECH AEC/APC Workshop IX, September 21-24, 1997.
5. A.J. Toprac, W.J. Campbell, 'APC Framework Run-to-Run Control Applications in Fab 25,' SEMATECH AEC/APC Symposium X, October 11-16, 1998.
6. W. Zhang, M. Richter, P. Solomon, Y. Kostoulas, G. Kneissi, W. Aarts, A. Waldhauer, 'Advanced Process Control for Epitaxial Silicon, ' Solid State Technology, September 1998.