Issue



PRODUCTIVITY & YIELD: Factory-wide run-to-run process control


12/01/1999







Mark Yelverton*, Advanced Micro Devices, Austin, Texas
Kostas Tsakalis, Arizona State University, Tempe, Arizona
Kevin Stoddard*, SEMY Engineering Inc., Phoenix, Arizona

Over the last several years, run-to-run process control has only been applied to select processes. Now, through advances in process-engineering-friendly software tools, it can be used across a wafer fab to maintain process repeatability automatically and compensate for upstream process variability, achieve better device yields and speeds, and greatly enhance factory productivity.

A common methodology for monitoring batch processes uses x-bar/s or x-bar
plots from statistical process control (SPC) software. Normally distributed process data is monitored using a set of rules (i.e., "Western Electric") to determine if a process is in control. Manual investigation and adjustment of the process are necessary when a data point is out of control. A large percentage of these adjustments are made to compensate for run-to-run variations attributed to process equipment drift.

Unfortunately, there are many problems with manually adjusted processes based on SPC charts. A typical wafer fab has ~2500 on-line SPC charts. If all Western Electric rules are used and if two new points are added to each chart/day, there could be an average 82 false alarms/day [1]. Only processes with the most significant excursions tend to be maintained due to the sheer magnitude of faults. In some cases, the opposite is true and too much attention is given to a chart and overadjustment occurs, resulting in processes "ringing." Additional process variation can be introduced between shifts or individuals as they try to compensate for each other's process adjustments (Fig. 1).

Automatic process control

Most problems associated with manual control of semiconductor processes can be eliminated with automatic process control. All areas in a wafer fab can show significant process control improvement after implementing even simple automated feedback process controllers [2, 3].

Click here to enlarge image

PHOTO. The screen shows a fault-detection chart from run-to-run process control software. The software automatically maintains repeatability for better device yields and factory productivity. (Computer illustration courtesy of SEMY Engineering)

There are two types of run-to-run process control — feedback and feed-forward. A feedback control system makes adjustments to recipe or process-tool parameters to maintain the desired end-of-run or in situ metrology results. Compensation for incoming post-process variations from previous steps is achieved using feed-forward process control; an open-loop relationship or model between process steps adjusts the process target in the current step.

Benefits from automatic run-to-run process control are numerous. Greater precision and accuracy are possible with smaller, more frequent adjustments. Control algorithms that ignore "flier" data points can be tuned for maximum and repeatable performance. Variables that once were considered too complex can be controlled. Human error can also be eliminated via a consistent adjustment methodology.

In contrast, manual adjustments are inherent approximations based on simple relationships, can be biased by past experience, or can be influenced by errors in reading or entering data. Once automatic control is implemented, personnel requirements to maintain a process are reduced, freeing engineers and operators to work on other issues.

Click here to enlarge image

Figure 1. SPC chart showing evidence of manual overadjustment of a CVD process deposition time.

Various obstacles must be overcome for successful implementation of a run-to-run control system. The automated acquisition of metrology results can, at times, be very difficult and must be extremely flexible. Some metrology tools do not provide proper connectivity to a factory MES system. Even in cases where metrology results are integrated into a centralized SPC database, custom interfaces are often required to communicate metrology data to the run-to-run control system. Moreover, metrology results must be acquired in a timely manner to be useful, especially in a feed-forward application

It is also necessary to implement proper fault detection and classification logic to deal with faulty metrology measurements caused by drifting metrology tools, operator error, or bad wafers. Proper classification of metrology data is essential to ensure that correct measurements are provided to the run-to-run controller. Further integration with the process tool (i.e., recipe management) is also required to provide the process adjustment mechanism for tools that do not directly support adjusting process parameters. This should include boundaries of adjustment and means to handle a control system failure.

Run-to-run process control

The implementation of a run-to-run process controller can be achieved in several ways. Direct implementation on a process tool allows for wafer-to-wafer process adjustments, if necessary, and ease of execution.

It does not allow feed-forward control across different processes, however, and requires costly dedicated or on-board metrology measurement.

Factory-wide implementation of run-to-run process control, such as Sematech's Advanced Process Control (APC)

Framework, is gaining acceptance in a few companies. Such a framework, which provides a communications bus and standards for building control and visualization modules, is extremely flexible, providing feedback and feed-forward functionality and the ability to exchange information between many suppliers. While the development and implementation of APC Framework modules are left entirely up to the user, the sheer magnitude and expense of such development may limit actual use in factories with a large diversity in products [4].

Run-to-run process control can also be implemented with off-the-shelf hardware and software, such as the Equipment Supervisor Workstation (ESW) developed by SEMY Engineering. This supervisory solution provides integration to virtually any process or metrology tool, and with SEMY's Advanced Run-to-Run Control (ARRC) module provides both feed-forward or feedback run-to-run process control (see sidebar "Advanced run-to-run control (ARRC) tools" on p. 46). Connectivity between ESW and APC Framework is designed around a Common Object Request Broker Architecture (CORBA). This solution is applicable to a single process tool, making it attractive for tool manufacturers or an entire process area [5]. An example of the latter is the implementation of this supervisory system at White Oak Semiconductor [6].

Process modeling and control

A typical methodology for implementing run-to-run process control uses a "black box" approach, where process-modeling and control-systems experts develop complex schemes to control the process accurately [6]. This approach is very powerful and can usually provide the best results; however, it can also be time-consuming to develop and implement, especially if a suitable control infrastructure is not in place. This infrastructure is provided by ARRC via "plug-in" algorithms using MATLAB — a high-performance numeric computation and visualization software package.

Click here to enlarge image

Figure 2. Oxidation process uniformity a) without and b) with run-to-run process control.

The "black box" approach does possess several drawbacks. Once developed, it is difficult or, in some cases, impossible to be modified or adjusted by production engineers. Also, current systems do not provide graphical user interfaces that allow easy gauging of the performance and accuracy of adjustments. ARRC, however, has a revolutionary set of tools that allows process engineers with little or no experience to develop their own models and control systems easily and analyze the performance of the control.

Feedback control examples

Simple feedback controllers can significantly improve process performance and productivity in every area of the production environment [7-9]. This method is useful for improving process control and automating routine adjustments made by operators, engineers, and maintenance personnel. The examples presented here are only a small part of what is achievable in a factory-wide implementation.

For example, in CMP operations, the polish time may be adjusted to control the remaining thickness of the film, and may be changed between wafers or between batches depending on the stability of the process. When wafer-to-wafer adjustments are required, in situ metrology is needed to provide measurements in time to close the loop. If the process drift is understood, a feedback model can be used to predict and adjust the polish time required for each wafer in a batch. The polish time can be changed from wafer to wafer based on model estimates, and then verified with metrology after the batch has been completed. Feedback controllers can also be used to compensate automatically for the changes in the slurry and degradation of the polish pads.

Diffusion processes often require the simultaneous adjustment of multiple variables. Low-pressure chemical vapor deposition (LPCVD) batch processes typically require temperature and time adjustments. A feedback controller can be used to adjust end-zone temperatures to minimize thickness differences between wafers processed in the center and end zones of the furnace. The feedback controller also adjusts deposition time to center the process at the desired thickness target.

Simple models are effective in both linear LPCVD deposition processes and nonlinear oxidation processes. In our work with oxidation processes, a process capability index (Cpk) improvement of 27% was achieved using a simple feedback controller (Fig. 2).

Feedback adjustments are useful in etch processes to control CDs. Many etch processes use in situ end-point detection. Once end point has been established, the recipe continues to etch the film for a pre-defined over-etch time. The impact of the end-point and over-etch times is measured in film thickness and CDs. Automatic feedback control can be applied to adjust timed etch processes or over-etch time in end-point driven processes. The relationship between film thickness removed and CDs to etch process parameters such as etch time, gas flow, and power can be modeled and controlled.

Feed-forward control examples

Although a process may be able to produce repeatable results using feedback control methodology, process results may also be dependent on the initial state of wafers. This information can be automatically provided through feed-forward modeling. However, prior to implementing any feed-forward technique, the process must be inherently stable or must use an effective feedback mechanism to provide stability.

One example where a feed-forward control mechanism is useful is the adjustment of etch time to remove an interlayer dielectric for a via interconnect. This type of etch process may not be controlled with in situ end-point detection, because the small amount of film being removed does not provide the needed signal strength. Instead, the process should remove all of the film in the first attempt, but the initial film thickness is required to select the target for the process.

In CMP processes, there are typical variations in initial surface material that result in similar variations after the polish. By measuring wafers prior to polish, a feed-forward controller can adjust the feedback controller target (amount of material to be polished) after each run to minimize or eliminate these variations.

Implant barrier variations can adversely affect the gain of a transistor. In this example, oxide and nitride layers are grown and deposited, respectively, on the wafer. Using lithography and etch processes, trenches are formed in the nitride layer. In forming the trench, the nitride is over-etched, resulting in removal of some of the initial oxide layer. A sacrificial oxide layer is grown in the trench over this initial oxide, making a barrier for the implant step. This implant barrier varies run-to-run due to the over-etch of the nitride and the variations of the initial oxide growth. The incoming variation caused by these etch steps can be minimized by first measuring the initial oxide layer after the nitride etch and adjusting the target of the feedback controller on the sacrificial oxide process to maintain a more consistent implant barrier (Fig. 3).

In photolithography, feed-forward control can be used to calculate alignment parameters from wafer to wafer. These are typically six to eight parameters that can be predicted with first-order models. Tilt can also be adjusted, but requires a much more complex model.

Conclusion

As factories look for new and innovative ways to reduce manufacturing costs, run-to-run process control solutions become increasingly important to squeeze the most performance out of processing tools. It has been shown that run-to-run process control provides significant process uniformity improvements, reduced process maintenance costs, and improved throughput, leading to a lower cost of manufacturing.

Click here to enlarge image

Figure 3. Feed-forward adjustment of sacrificial oxide growth to reduce implant barrier variation attributed to initial oxidation and etch process variation.

Our development and use of the ARRC system shows that it provides a seamless architecture for factory-wide run-to-run process control. Its tools allow users with little or no modeling and control expertise to create input-output relationships for any process and control them to a desired target. System flexibility allows users to implement more complex modeling and control methodologies using "plug-in" modules.

We have presented here a few example processes in CMP, diffusion, etch, and photolithography that can benefit from feed-forward and feedback process control, but applications are limited only by the imagination of the process engineer.

Acknowledgments

Additional authors include Mike Simpson and Brian Cusson of Advanced Micro Devices, and Pradeep Swamy, Brad Schulze, and Kevin Dimond of SEMY Engineering. This project was sponsored in part by Sematech's Equipment Productivity Improvement Team program. The authors thank Abhijit Bora, Tony Colombo, Keith Laidlaw, Tarig Mutlag, Scott Say, Ashok Tripathi, and Renran Yiu of SEMY Engineering, and Tom Timmons of Advanced Micro Devices for their invaluable contributions. MATLAB is a registered trademark of The MathWorks Inc. Equipment Supervisor Workstation and Advanced Run-to-Run Control are registered trademarks of SEMY Engineering.

References

  1. R. Patty, "A More Robust and Reliable Statistical Process Control System," International Symposium on Semiconductor Manufacturing, 1999.
  2. K. Stoddard et al., "Application of Feed-Forward and Adaptive Feedback Control to Semiconductor Device Manufacturing," American Control Conference, Baltimore, Maryland, June 1994.
  3. N. Zhe et al., "A Comparative Analysis of Run-to-Run Control Algorithms in the Semiconductor Manufacturing Industry," IEEE/SEMI 1996 Advanced Semiconductor Manufacturing Conference & Workshop, pp. 375-381, New York, 1996.
  4. M. Miller, "From APC Pilot to Full Production System — Scaling Up is Hard to Do," AEC/APC Symposium XI, Vail, Colorado, September 1999.
  5. M. Yelverton et al., "A Complete Furnace Control Platform for High-Volume Manufacturing," International Symposium on Semiconductor Manufacturing, Tokyo, Japan, October 1998.
  6. T. Dowd, "Results of AEC/APC Deployment at White Oak Semiconductor," AEC/APC Symposium XI, Vail, Colorado, September 1999.
  7. M. Hankinson et al., "Integrated Real-Time and Run-to-Run Control of Etch Depth in Reactive Ion Etching," IEEE Transactions Semiconductor Manufacturing, Vol. 10, No. 1, pp. 121-130, February 1997.
  8. T.H. Smith et al., "Run-by-Run Advanced Process Control of Metal Sputter Deposition," IEEE Transactions Semiconductor Manufacturing, Vol. 11, No. 2, pp. 276-284, May 1998.
  9. S.F. Lee, C.J. Spanos, "Prediction of Wafer State After Plasma Processing Using Real-Time Tool Data," IEEE Transactions Semiconductor Manufacturing, Vol. 8, No. 3, August 1995.

For more information, contact Kevin Stoddard at SEMY Engineering Inc., 2340 West Shangri La Rd., Phoenix, AZ 85029; ph 602/861-9395, fax 602/861-1442, e-mail [email protected].


Advanced run-to-run control (AARC) tools

ARRC's run-to-run feedback adjustments (shown in blue, Fig. 1) can originate with metrology data from a MES system, more efficient direct data from metrology tools, or data from an APC Framework. The system ensures incoming data fall within guidelines. It then combines the data to define control points used by an algorithm to generate process-variable adjustments within the process recipe or through tables or parameters that are downloaded directly to the tool.

Click here to enlarge image

Figure 1. Run-to-run process control from feedback (blue) and feed-forward (red) components.

ARRC's feed-forward process control (shown in red, Fig. 1), which uses metrology acquisition, manipulation, and models common to feedback control, is achieved using open-loop control to adjust algorithm targets for the current process based on upstream process measurements.

Depending on the metrology data and the process adjustment mechanism, ARRC makes process adjustments at varying frequencies. For batch tools, it adjusts parameters on a batch-to-batch basis; for single-wafer tools, it can adjust parameters lot-to-lot or wafer-to-wafer.

"Process-engineering-friendly" tools include an empirical model builder that characterizes relationships between process measurements and control variables. Users enter a "golden" set of input-output data obtained from first principles, experimental results, or a design of experiment, defining a model type (i.e., single-input single-output or multi-input multi-output polynomials of various orders) that best fits the shape or structure of the data. ARRC creates and displays the model against input-output data, providing process noise and model integrity metrics to quantify the model. Model parameters are computed using a simple, reliable, least-squares estimation. The user can iteratively choose different model types to obtain the best fit.

Automatic on-line adaptive modeling is done on a run-to-run basis using the latest process results to compensate for process tool drift (Fig. 2). Users only need modest modeling expertise. Adaptive modeling uses a modified fading-memory, least-squares algorithm incorporating dead-zones and parameter constraints. The modifications, together with the ability to perform full or partial adjustment of the parameters, aim to alleviate potential problems due to estimation drift. This can occur when the input-output data used for estimation do not provide sufficient information to determine all model parameters (e.g., normal production data).

Click here to enlarge image

Figure 2. Improvements achieved measured as a) process average and b) process range without manual run-to-run process adjustment, and with recursive and adaptive control.

The basic control algorithm is of the gradient-Newton class used to solve nonlinear equations. These have excellent convergence properties near the solution, but may be susceptible to noise. Thus, ARRC modifies the algorithm with a "dead-zone," nonlinear gain adjustment allowing users to account for sensitivity to noise. This allows for a quick response to large disturbances without increasing the steady-state variance [1, 2]. Higher-order controllers can also be used to compensate for more severe process drifts. Default control parameters are automatically computed based on the model and the input-output data set. These parameters are fully user-adjustable to refine the control action.

A process-metrology tool simulator is also provided to further assist the user in visualizing operation of the feedback controller and learning how to adjust the control parameters to achieve the desired behavior before performing actual process test runs.

References

  1. K. Tsakalis et al., "Optimizing Diffusion Furnace Performance Using Run-to-Run Control," Semiconductor Process Analysis & Control Conference, San Antonio, Texas, January 1999.
  2. K. Tsakalis, "The Role of Dead-Zones in Improving Run-to-Run Control Performance," AEC/APC Symposium XI, Vail, Colorado, September 1999.