ADVANCED process control (APC) has been recognized

Transcription

1 276 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 11, NO. 2, MAY 1998 Run by Run Advanced Process Control of Metal Sputter Deposition Taber H. Smith, Duane S. Boning, Member, IEEE, Jerry Stefani, Member, IEEE, and Stephanie Watts Butler Abstract Metal sputter deposition processes for semiconductor manufacturing are characterized by a decrease in deposition rate from run to run as the sputter target degrades. The goal is to maintain a desired deposition thickness from wafer to wafer and lot to lot. Run by run (RbR) model-based process control (MBPC) has been applied to metal sputter deposition processes at Texas Instruments. RbR MBPC, based on the exponentiallyweighted moving-average filter, provides the ability to track and compensate for process drifts without a priori assumptions on their magnitude or consistency (from sputter target to sputter target or collimator to collimator). The application of RbR MBPC resulted in an improved C pk of 44% for aluminum sputter deposition, while reducing the number of lot-based monitor wafers by a factor of three. The application of RbR MBPC to the titanium sputter deposition process eliminated look-ahead test runs and reduced the number of monitor wafers by a factor of three. At the same time, C pk was improved by 10% with the application of RbR MBPC. Index Terms Advanced process control, exponentially weighted moving average, predictor corrector control, run by run process control, sputter deposition. I. INTRODUCTION ADVANCED process control (APC) has been recognized as an enabling technology for meeting the increased efficiency and product quality demands which will drive the future profitability of semiconductor manufacturing facilities. The building blocks of an advanced control system are industrial-quality APC methods, sensor and diagnostic technologies, and integration tools. The introduction of new sensors and process diagnostic tools, the maturation of existing technologies in these areas, and the development of integration tools and methodologies are making the application of APC methodologies a reality for the industry. Currently, there exists three primary focus areas for the application of APC techniques: statistical process control (SPC) with model-based process control (MBPC) [1] [3], run by run (RbR) MBPC [4] [10], and real-time process control (RTPC) [11] [15]. The application of RTPC is often difficult due to current limitations of integration with tool software and reliability of, and longterm reproducibility of, real-time sensors [16]. On the other hand, SPC with MBPC and RbR MBPC are in the final Manuscript received May 15, 1997; revised October 28, T. H. Smith and D. S. Boning are with the Microsystems Technology Laboratories, Massachusetts Institute of Technology, Cambridge MA USA ( taber@mit.edu; boning@mtl.mit.edu). J. Stefani and S. W. Butler are with the Semiconductor Process and Device Center, Texas Instruments Incorporated, Dallas, TX USA ( stefani@spdc.ti.com). Publisher Item Identifier S (98) development stages, with experimental results demonstrating significant improvement in wafer processing efficiency and quality (in terms of test and monitor wafer savings, reduced machine downtime, increased wafer throughput, increased process capability, and decreased product scrap) [17]. RbR MBPC methods are most applicable to processes which exhibit good quality and reliability within a lot, but whose lot to lot (run by run) performance is plagued by machine wear or inaccurate assumptions made about the consistency of replaceable components. These characteristics suggest that control actions made on a run by run basis may be helpful. In addition, many tools have outputs whose values are affected by machine wear across all processes performed in that tool, but at the same time these values are dependent on the particulars of each process. The use of a process model in MBPC provides the ability to incorporate these process-dependent effects while maintaining an accurate estimate of how the machine state changes from lot to lot. MBPC often results in reduced wafer processing complexity, increased processing efficiency, and improved processing quality. The reduced complexity arises because the controller separates the process and machine dependencies, thereby allowing the operators the freedom from monitoring these differences themselves. The increased efficiency is due to the decreased number of lookahead and lot-based monitor wafers required to effectively control a process. The separation of process variation from the machine variation allows the controller to better estimate the machine state and utilize this information to increase the product quality. Increased product quality is associated with improved process performance, such as process capability, or Metal sputter deposition for semiconductor manufacturing, as we will show, can benefit greatly from RbR MBPC. In particular, sputter deposition experiences a characteristic drift in the deposition rate which occurs across all processes in the tool (or a single chamber of the tool). There is also a dependency of the deposition rate on the final film thickness (process dependence). In this paper, we will demonstrate that these characteristics make RbR MBPC an effective control method for metal sputter deposition. We will apply two RbR MBPC methods, the EWMA controller and the predictorcorrector controller, to two metal sputter deposition processes. We will first employ these controllers in a focused application to the aluminum sputter deposition process with a single thickness goal. We will then describe the general application of these controllers to sputter deposition processes. For titanium, we will use a general modeling approach to separate the /98$ IEEE

2 SMITH et al.: ADVANCED PROCESS CONTROL OF METAL SPUTTER DEPOSITION 277 Fig. 2. Historical data for aluminum sputter deposition rate. Fig. 1. Diagram of metal sputter deposition process. process dependency (rate versus thickness) from the equipment disturbances (drift and noise). We will also compare the two RbR MBPC techniques and identify where each is strongest. Section II describes the sputter deposition process and examines historical data for two specific sputter deposition processes. The run by run methods and models which characterize the sputter deposition processes are presented in Section III. Section IV contains the experimental results. Finally, Section V presents conclusions and discusses future work. II. SPUTTER DEPOSITION PROCESSES A basic diagram for metal sputter deposition is shown in Fig. 1. Ions are discharged from the plasma which dislodge particles from the metal sputter target. These particles condense on the surface of the wafer. A honeycomb-like collimator is often placed above the wafer to improve bottom coverage in contacts/vias. The goal is to maintain a desired deposition thickness from wafer to wafer and lot to lot. Achieving the desired film thickness can be difficult because metal sputter deposition processes are characterized by a decrease in deposition rate as the sputter target degrades and material builds up in the collimator. Inconsistencies in the gas flows, vacuum pressure, supplied power, and molecular uniformity of the sputter target contribute various amounts of noise, or variation, in the measured deposition rate in addition to the characteristic drift which occurs from wafer to wafer. In addition, a single chamber in a sputter tool may process several different lot types in which the desired film thicknesses can span a wide range of values. A variety of process goals adds another level of difficulty because the deposition rate can depend on the final film thickness. In this paper, we consider two metal sputter deposition processes, aluminum and titanium, on an Applied Materials Endura 5500, in the Semiconductor Process and Device Center (SPDC) at Texas Instruments. A. TiN/Al-0.5% Cu/TiN Stack The first process we considered was the deposition of aluminum with 0.5% copper at a single desired thickness in a stack of titanium-nitride (TiN), Al-0.5% Cu, and TiN, without the use of a collimator. The TiN thicknesses are small relative to the aluminum thickness and are not considered to adversely affect the measurement of the aluminum sheet resistance. Consider the historical data for this process shown in Fig. 2. Here we can see the characteristic drift in the deposition rate caused by the degradation of the target over the long term (in kilowatt-hours, KWH). It can also be seen from Fig. 2 that the rate at which the process drifts varies from target to target, and the drift rate may change over the life of a single target. Finally, the starting deposition rate may differ from target to target, as well. B. Titanium The second process we considered was the collimated deposition of titanium with desired deposition thicknesses spanning a range of 400 Å. As with the aluminum case, the deposition of titanium is characterized by a persistent drift with significant process noise. In fact, the drift in deposition rate is more pronounced for titanium (larger percentage decrease) due to the build up of material inside the collimator. Fig. 3(a) shows the histories for two consecutive collimators for a range of thicknesses as a function of the collimator age (in KWH). The rate appears quite noisy. However, for a reduced range of thicknesses for collimator #2, shown in Fig. 3(b), the rate appears less noisy, and the persistent drift is more apparent. Similar results were found for small ranges of thicknesses centered elsewhere across the 400-Å total range. These data suggest that the deposition rate is a function of the final film thickness. III. RUN BY RUN MODEL-BASED PROCESS CONTROL Maintaining a desired film thickness for the processes described in Section II is difficult since we cannot know a priori what the magnitude of the drift will be. Therefore, in order to control the process, we must maintain an accurate estimate

3 278 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 11, NO. 2, MAY 1998 (a) Controlled thickness for MBPC with SPC for metal sputter deposi- Fig. 4. tion. (b) Fig. 3. (a) Titanium sputter deposition rate versus collimator age for all thicknesses. (b) Titanium sputter deposition rate versus collimator age for a 100 Å range of final film thicknesses. of the deposition rate over the life of the sputter target and collimator. Two common methods for control assume the drift rate is constant over the life of a sputter target. These schemes are typically employed internally to the sputter deposition tool itself. Open-loop compensation strategies which make this constant drift assumption assume the rate on run n is a function of the form Rate Rate (1) where is a fixed (negative) constant and is the age of the sputter target in KWH. In the first control method, termed time compensation, the sputter tool utilizes (1) to compensate for the drop in rate by incrementally adding additional process time to the recipe to maintain the thickness goal over the long term. The engineer specifies the value in (1) for each new target/collimator. A second method, termed power compensation, also assumes a constant drift in rate while adding the power setting to the underlying rate function as follows: Rate Rate (2) where is a known function of the power,, on run The sputter tool compensates for the decrease in rate due to the drift term by increasing power to maintain Rate Rate Again, the engineer specifies explicitly the increase in watts per KWH through While both time and power compensation can be effective control strategies, they inherently do not take into account that each target/collimator generally does not have the same drift rate or that the rate may change over the life of the target/collimator. Thus, these open-loop control techniques require additional supervision in the form of continual monitoring and tweaking of the process time (whether in an ad hoc fashion or utilizing, for instance, SPC techniques) to maintain the deposition thickness at a desired value. This supervision typically requires many test runs and/or lot-based monitor wafers, increasing cost and decreasing the time the tool is available to do useful work. For those tools which are not capable of internal compensation (either time or power), an alternative combines SPC with MBPC. In SPC with MBPC, SPC is used to monitor the drift in deposition rate. When control limits are reached, or when some supplementary statistical control rules are violated, the rate estimate (the process model) is updated. The update generally includes calculating a new rate based on the data points which triggered the out-of-control condition. The updated rate estimate is used to calculate new process times for subsequent lots. An example of the controlled thickness resulting from SPC with MBPC with control limits only is shown in Fig. 4. Fig. 4 indicates that while reasonable control is achieved, there exists room for further improvement between model updates. One could argue that a better choice of control limits or statistical rules could improve process performance. However, SPC does not take advantage of our knowledge that the process is, in fact, drifting. RbR MBPC methods, though, are excellent for controlling drifting processes. A. Controllers Based on Exponentially-Weighted Moving-Average Filtering The drifting nature of metal sputter deposition indicates that continual changes to the process model and recipe are

4 SMITH et al.: ADVANCED PROCESS CONTROL OF METAL SPUTTER DEPOSITION 279 Fig. 5. Example of EWMA rate estimation and PCC rate prediction for a drifting process. necessary to compensate for the drift. A more appropriate model for metal sputter deposition is of the form Rate Rate (3) where Rate is the estimate of the rate at the beginning of the target/collimator life and is the estimated error in the deposition rate (from Rate[0]) on run In addition, we would like to adjust this estimate and make changes to the deposition time based on (3) every run (or lot). In this paper, we consider two particular implementations of this strategy, referred to as RbR MBPC. These methods are based on the exponentially-weighted moving-average (EWMA) filter and have been shown to provide excellent control of processes with a persistent drift [8]. The first implementation of a RbR MBPC strategy, referred to here as the EWMA controller, uses a single EWMA filter to update the value of every run (based on a lot-based monitor wafer). This update is of the form: where Equivalently, Rate Rate (4) Rate Rate (5) Rate Rate is the filter factor, or weight. The time for the next run is then computed using (3) as follows: Desired Thickness Process Time (7) Rate The EWMA controller effectively smoothes out the past data every run in order to best estimate the current process state. The EWMA controller is shown schematically in Fig. 5 for a drifting process. A weighted average of the model errors from previous runs is used to update the current estimate of the rate. If the drift is relatively slow, with frequent model (6) Fig. 6. Deposition rate versus final film thickness for titanium sputter deposition. updates, the EWMA controller does a good job of tracking the process state. For instance, the predicted deposition rate for run number 35 in Fig. 5 using the errors from runs 34, 33, etc., is fairly close to the actual value of the rate for run 34. Thus, with a small drift, this estimate will be accurate for run 35. However, for processes with a relatively large drift and/or infrequent model updates, the EWMA controller alone is often not sufficient. In particular, the process at run 35 has drifted further away from the predicted process state (the rate in run 34). As a result, the controller model is again offset from the true rate. This lag is characteristic of a controller trying to compensate a drifting process using past errors to estimate larger future errors. This lag results in a nonzero average deviation between the actual output and the goal. While the EWMA controller often provides excellent control of slowly drifting or wandering noise (ARMA) processes, a more sophisticated controller is required for larger drifts. An extension of the EWMA controller, the predictor-corrector controller (PCC), incorporates two EWMA filters to tune the process model [7]. In particular, PCC adds an EWMA estimate of the prediction error in the EWMA controller (also called the trend in the error). The prediction of the rate for the next run is then based on the estimate plus the estimate of the trend in the error, Trend This is summarized by the following equations: Trend Rate Rate (8) Rate Trend Rate Trend Trend (9) where Rate Trend is often taken to be the following: RateTrend Rate Rate (10) For comparison, the PCC is shown schematically in Fig. 5, where the EWMA estimated prediction error, or trend, is added to the EWMA estimated rate. This method allows the controller to remove the lag in the EWMA controller.

5 280 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 11, NO. 2, MAY 1998 Fig. 7. Deposition rate versus final film thickness data at the target/collimator ages in Table I. B. Utilizing EWMA-Based Controllers to Accommodate More Complex Models For aluminum, (3) applies across a range of thickness goals. We can determine a starting value for Rate from historical data. However, (3) does not apply directly for the titanium deposition process. For titanium, the deposition rate is a function of the deposited film thickness. 1 Fig. 6 shows an example of the dependence of rate on thickness for the titanium process with a new sputter target and new collimator. The apparent increase in deposition rate with increasing film thickness is in fact an increase in deposition rate with increasing process time due to process start-up effects. We would like to describe the dependence in Fig. 6 by the following relationship: Rate Desired Thickness Rate Desired Thickness (11) where Rate Desired Thickness is a simple polynomial expression representing the solid line in Fig. 6. Equation (11) says that the dependence of rate on thickness is invariant with respect to overall decreases in the deposition rate. Thus, the deposition rate can be determined for any thickness goal at any time during the life of the target/collimator by simply estimating a single value, Equation (11) also implies that test data from only a single thickness, not multiple thicknesses, is needed at any process state to update If (11) is valid, we can save a significant number of test wafers. Equation (11) is valid only if the dependency of rate on thickness does not change as the target/collimator ages. To prove (11) for titanium sputter deposition, experiments were performed at different combinations of the ages of the sputter target and collimator. The test points are shown in Table I. The experimental results are shown in Fig. 7. Fig. 7 reveals that although the curves are offset from one another (due 1 Deposition rate is defined as the average film thickness divided by process time. Note that this rate is not equal to the instantaneous rate d(thickness)/dt, unless the instantaneous rate is constant over all thicknesses. TABLE I EXPERIMENTAL MATRIX FOR VALIDATION OF (11) FOR TITANIUM SPUTTER DEPOSITION to different target/collimator ages), the higher order terms (shapes of the curves) are relatively constant. Thus, we can utilize (11) for titanium together with the EWMA and PCC controllers to try to achieve reduced processing complexity and increased processing efficiency. Specifically, we can update the process model for all thickness goals using the most recent measurement, regardless of the exact film thickness. This greatly reduces the number of required test wafers. IV. EXPERIMENTAL RESULTS AND DISCUSSION The EWMA and PCC controllers were applied to the aluminum and titanium deposition processes in the Semiconductor Process and Device Center at Texas Instruments. The aluminum experiment covered the life of one target and two process kits, while the titanium experiment covered three different sputter targets and four collimator changes. The experimental results are highlighted below and compared to the corresponding existing control practices. A. TiN/Al-0.5% Cu/TiN Stack Results The EWMA and PCC controllers were applied to aluminum and titanium processes according to Fig. 8. In particular, for each lot the process time (recipe) for the desired film thickness is first downloaded from the controller to the tool. In all cases the power setting was fixed. The lot is then processed, and the monitor wafer is measured. The wafer-state responses (sheet resistances) are next uploaded to the controller. The updated process model is used to calculate the process time

6 SMITH et al.: ADVANCED PROCESS CONTROL OF METAL SPUTTER DEPOSITION 281 Fig. 8. Data flow in the RbR MBPC system for metal sputter deposition process. Fig. 10. Distribution of errors from desired thickness (%) for aluminum sputter deposition with the EWMA controller. Fig. 9. State estimation results for EWMA controller applied to aluminum metal sputter deposition. The solid line is the measured deposition rate. The dashed line is the smoothed rate estimate. for the subsequent lot. The measured rate and corresponding estimate of the process state for the aluminum experiment is given in Fig. 9. The EWMA controller accurately tracked the decrease in deposition rate over the life of the sputter target. The distribution of errors from the desired thickness for the EWMA controller is revealed in Fig. 10. With RbR MBPC, control of aluminum thickness was to within 3% of the goal, compared to approximately 5% without MBPC (Fig. 11). the process capability, was improved by 44% with the EWMA controller. The PCC controller was also applied to the aluminum sputter process. The distribution of errors for the PCC are plotted in Fig. 12. improved by 39% with PCC compared to before RbR MBPC. In this case, the addition of the EWMA controller prediction error removed the slight offset in the mean of the distribution. However, the process variance (spread of the distribution) slightly increased with PCC compared to the EWMA controller. The better overall performance of the EWMA controller was mainly due to timely process sampling, poor initial estimates of the EWMA weights for PCC, and the relatively noisy aluminum metal sputter process (making it difficult to estimate the trend in the error). The choice of weights for both RbR controllers is critical for achieving Fig. 11. Distribution of errors from desired thickness (%) for aluminum sputter deposition before RbR MBPC. optimal performance. In-depth discussions regarding optimal weight determination are given elsewhere [18]. In addition to the improved performance achieved by these RbR MBPC methods, the number of monitor wafers was reduced from one in each lot without RbR MBPC to one in every three lots with RbR MBPC. At the same time, lookahead test wafers were eliminated. B. Titanium Results Current practice in SPDC for titanium deposition is very different than for TiN/Al-0.5% Cu/TiN. As opposed to a

7 282 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 11, NO. 2, MAY 1998 Fig. 14. Error from desired thickness (%) results for EWMA controller for titanium sputter deposition. Fig. 12. Distribution of errors from desired thickness (%) for aluminum sputter deposition with the PCC. Fig. 15. State estimation results for PCC applied to titanium sputter deposition. The solid line is the measured deposition rate. The dashed line is the smoothed rate estimate. Fig. 13. State estimation results for EWMA controller applied to titanium sputter deposition. The solid line is the measured deposition rate. The dashed line is the smoothed rate estimate. single thickness goal for aluminum, more than five different thicknesses of titanium are deposited on a regular basis. The current practice is to run lot-based monitor wafers for all lots. Monitor wafers for one thickness are used to estimate rates for other thicknesses within 100 Å. However, lots with a thickness goal which has not been run recently are preceded by a look-ahead test wafer. Due to the many monitor and look-ahead wafers, control of titanium sputter deposition before RbR MBPC was quite good. As a result, the goal for RbR MBPC for titanium was increased processing efficiency through the reduction of monitor and look-ahead wafers, while maintaining or improving process performance. As described earlier, processing was simplified with RbR MBPC by allowing data for any film thickness to be used to update the process model. This eliminated the need for look- ahead wafers prior to lot processing (previously approximately one every three lots). The number of monitor wafers during lot processing was reduced from one every lot to one every three lots. To challenge the RbR controllers, lot-based monitors were intentionally run at a variety of thicknesses in a cyclic manner (e.g., 100, 200, 300, 200, 100 Å, etc.) This insured that each film thickness in the range was used to control the process and was used at about the same frequency. The state estimation results for the EWMA controller applied to titanium sputter deposition are presented in Fig. 13. Because of the relatively small noise to drift ratio, the offset inherent in the EWMA controller explained earlier is, in this case, readily apparent. The effect is also apparent in the corresponding film thickness control results plotted in Fig. 14. Due to the lag, process performance actually decreased compared to before RbR MBPC. However, whereas for aluminum sputter deposition we did not see a significant impact, the benefits of PCC for titanium sputter deposition are clear (Fig. 15). The PCC effectively removed the offset in

8 SMITH et al.: ADVANCED PROCESS CONTROL OF METAL SPUTTER DEPOSITION 283 becomes imperative. Therefore, future work will focus on the implementation of a self-tuning EWMA controller such as that described in [18]. In addition, more complex processes may require the use of general approximation structures for representative process models for control. A second goal for the future will be the implementation of general internal model control (IMC) architectures such as that described in [7], [19]. Finally, the need for control methods which facilitate model updates at infrequent and nonconstant sampling periods, while still containing feed-forward estimates of the process state, will provide the potential for further reduction of test wafers. Fig. 16. Error from desired thickness (%) results for PCC for titanium sputter deposition. Fig. 13. Even with relatively infrequent updates to the process model, the PCC controller accurately tracked the change in deposition rate over the life of the collimators and sputter targets. The corresponding film thickness control results are given in Fig. 16. The PCC improved by 10% compared to before RbR MBPC, while eliminating look-ahead tests and reducing the number of monitor wafers by a factor of three. V. CONCLUSIONS AND FUTURE WORK RbR MBPC has been applied to metal sputter deposition processes for semiconductor manufacturing. RbR MBPC provides the ability to track and compensate for process drifts without a priori assumptions on their magnitude or consistency (from target to target or collimator to collimator). For processes with a relatively large amount of noise relative to the magnitude of the drift, namely aluminum sputter deposition, we have demonstrated that the EWMA controller is very effective in improving process capability and reducing test wafers. The EWMA controller, though, falls short in the presence of a relatively large drift to noise ratio (i.e., for titanium). In this case, PCC provides a more accurate rate estimator. We have outlined how both RbR controllers can be modified to allow more complex process models to be incorporated, especially to accommodate the dependency of deposition rate on metal film thickness. The application of RbR MBPC resulted here in significantly improved control of the aluminum sputter deposition process, by 44%, while reducing the number of monitor wafers by a factor of three. The application of RbR MBPC to the titanium sputter deposition process eliminated look-ahead tests and reduced the number of monitor wafers by a factor of three. At the same time, process capability for titanium sputter deposition improved by 10% with the application of RbR MBPC. The implementation of RbR MBPC methods requires a discriminate method for properly selecting optimal filtering weights. As increasingly more complex process control algorithms make their way into semiconductor manufacturing, the need for turn-key control solutions for process engineers REFERENCES [1] P. K. Mozumder and G. G. Barna, Statistical feedback control of a plasma etch process, IEEE Trans. Semiconduct. Manufact., vol. 7, pp. 1 11, Feb [2] P. K. Mozumder, S. Saxena, and D. J. Collins, A monitor wafer based controller for semiconductor processes, IEEE Trans. Semiconduct. Manufact., vol. 7, pp , Aug [3] J. Stefani, S. Poarch, S. Saxena, and P. K. Mozumder, Advanced process control of a CVD tungsten reactor, IEEE Trans. Semiconduct. Manufact., vol. 9, pp , Aug [4] E. Sachs, A. Hu, and A. Ingolfsson, Run by run process control: Combining SPC and feedback control, IEEE Trans. Semiconduct. Manufact., vol. 8, pp , Feb [5] S. Leang, S.-Y. Ma, J. Thompson, B. J. Bombay, and C. J. Spanos, A control system for photolithographic sequences, IEEE Trans. Semiconduct. Manufact., vol. 9, pp , May [6] G. Box and T. Kramer, Statistical process monitoring and feedback adjustment A discussion, Technometrics, vol. 34, no. 3, pp , [7] S. W. Butler and J. Stefani, Supervisory run-by-run control of polysilicon gate etch using in situ ellipsometry, IEEE Trans. Semiconduct. Manufact., vol. 7, pp , May [8] T. Smith, D. Boning, J. Moyne, A. Hurwitz, and J. Curry, Compensating for CMP pad wear using run by run feedback control, in Proc. VLSI Multilevep Interconnect Conf., 1996, pp [9] M. Le, T. Smith, D. Boning, and H. Sawin, Run to run model based process control on a dual coil transformer coupled plasma reactor, in Proc. 191st Meeting Electrochem. Soc., [10] D. Boning, W. Moyne, T. Smith, J. Moyne, and A. Hurwitz, Practical issues in run by run process control, in Proc. IEEE/SEMI Advanced Semiconductor Manufacturing Conf., 1995, pp [11] B. Rashap et al., Control of semiconductor manufacturing equipment: Real-time feedback control of a reactive ion etcher, IEEE Trans. Semiconduct. Manufact., vol. 8, pp , Aug [12] T. E. Benson, et al., Sensor systems for real-time feedback control (RTC) of reactive ion etching, J. Vac. Sci. Technol. B, vol. 14, no. 1, pp , [13] M. Chandhok, P. D. Hanish, and J. W. Grizzle, Phenomenological modeling for real-time feedback control of RTE, in Proc. Process Control, Diagnostics, & Modeling in Semiconductor Manufacturing Symp., The Electrochemical Society, vol. 95, no. 2, pp , [14] T. F. Edgar, T. Breedijk, S. G. Bushman, S. W. Butler, S. Chatterjee, A. J. Toprac, and I. Trachtenberg, Modeling and control of microelectronics materials processing, Comput. Chem. Eng., vol. 19, no. 1, [15] R. Telfeyan, J. Moyne, N. Chaudhry, J. Pugmire, S. Shellman, D. Boning, W. Moyne, A. Hurwitz, and J. Taylor, A multi-level approach to the control of a chemical-mechanical planarization process, J. Vac. Sci. Tech. A, vol. 14, no. 3, pp , [16] G. Barna, et al., Sensor integration into plasma etch reactors of a developmental pilot line, J. Vac. Sci. Tech. B, vol. 12, no. 4, pp , [17] S. Muthukrishnan and J. Stefani, SCFab model-based process control methodology: Development and deployment for manufacturing excellence, TI Tech. J., vol. 13, no. 5, pp. 9 15, [18] T. Smith and D. Boning, A self-tuning EWMA controller utilizing artificial neural network function approximation techniques, IEEE Trans. Comp., Packag., Manufact. Technol. C, vol. 20, pp , Apr [19] T. Smith and D. Boning, An artificial neural network EWMA run by run controller for semiconductor processes, J. Vac. Sci. Technol. A, vol. 15, no. 3, pp , 1997.

9 284 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 11, NO. 2, MAY 1998 Kappa Phi. Taber H. Smith received the B.S. degree (with highest honors) from the Rochester Institute of Technology, Rochester, NY, in He received the M.S. degree from the Massachusetts Institute of Technology in 1996, where he is currently pursuing the Ph.D. degree. His principle research interests include distributed computing, process modeling, as well as the application of statistical and intelligent control methods to the semiconductor industry. Mr. Smith is a member of Tau Beta Pi and Phi Duane S. Boning (M 91) received the S.B. degrees in electrical engineering and in computer science in 1984, and the S.M. and Ph.D. degrees in 1986 and 1991, respectively, all from the Massachusetts Institute of Technology (MIT), Cambridge. He was an NSF Fellow from 1984 to 1989, and an Intel Graduate Fellow in From 1991 to 1993, he was a Member of the Technical Staff at the Texas Instruments Semiconductor Process and Device Center, Dallas, TX, where he worked on process/device simulation tool integration, semiconductor process representation, and statistical modeling and optimization. He is currently an Associate Professor of Electrical Engineering and Computer Science at MIT. His research focuses on variation modeling and control in semiconductor processes, with special emphasis on chemical-mechanical polishing and plasma etch. Additional interests include tools and frameworks for process and device design, network technology for distributed design and fabrication, and computer integrated manufacturing. Dr. Boning is an Associate Editor for the IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, and is a member of Eta Kappa Nu, Tau Beta Pi, Sigma Xi, and the Association of Computing Machinery. Jerry Stefani (M 90) received the M.S. and Ph.D. degrees in metallurgy/materials science at Columbia University, New York, NY, in 1985 and 1988, respectively. He received the B.S. degree in mechanical/materials engineering from the University of California, Davis, in Since 1989, he has worked in the Semiconductor Process and Device Center, Texas Instruments, Dallas, TX. He is presently a Senior Member of the Technical Staff in the advanced process control group. He applies statistical techniques to semiconductor manufacturing equipment with the overall goal of improving product yield. These techniques include statistical design of experiments, statistical process control, and advanced model-based process control. He recently participated in the ARPA-sponsored Microelectronics Manufacturing Science and Technology (MMST) project at TI where he developed advanced process control tools for integrated circuit fabrication. He is presently working on the productization and deployment of TI s ProcessWORKS TM advanced process control system. Stephanie Watts Butler received the B.S. degree in chemical engineering from Oklahoma State University, Stillwater, and the Ph.D. degree in chemical engineering from the University of Texas at Austin. She is currently the manager of the Advanced Process Control Branch, Semiconductor Process and Device Center, Texas Instruments, Dallas, TX. Previously, she worked on the successful MMST program developing control strategies and robust endpointing algorithms. She also has process development experience in lithography, wet clean, etching, and silicon crystal growth at both Texas Instruments and IBM. She is a Research Affiliate at the Massachusetts Institute of Technology, Cambridge, and past Co-Chair of the SEMATECH Advanced Process & Equipment Control FTAB. She is an author of over 20 papers and presentations and several patents.