Neal Lafferty, Rochester Institute of Technology; and Bennie Fiol, Paul Jowett, Yuri Karzhavin, and Tim Urenda, Infineon

Experiments performed during implementation of a real-time process control technique revealed the technique's advantages over traditional equipment qualification.

With the high cost of IC manufacturing equipment and an increasingly competitive memory chip market, it is imperative that fabs utilize existing manufacturing resources to their fullest capabilities. Substantial increases in productivity are possible by monitoring critical process parameters in real time using data from product wafers. Modern data acquisition systems in combination with in-line metrology tools make this advanced process control (APC) concept possible, while the larger number of devices on 300-mm wafers, and therefore their greater value compared with their 200-mm counterparts, will necessitate APC's adoption.¹

The primary goals of APC are to increase yields and decrease manufacturing costs. Figure 1 compares the monetary losses associated with various strategies used to monitor a typical etch process. As the strategies' mean time to detect (MTTD) an out-of-control (OOC) process increases, the associated loss in dollars increases exponentially, since additional processes are performed on defective, low-yielding wafers. Most fabs use an in-line, nonintegrated measurement strategy to monitor process performance. In that scenario, several lots will be processed on a tool while measurements of a previously processed lot are being obtained. If dynamic sampling is used, the number of lots processed before detection may be even higher. Thus, many lots could be affected by a single OOC excursion. In the next-best case, an integrated measurement configuration, each lot is sampled immediately after etching. In this case, only a few lots will be processed before measurements indicate that a problem exists; however, entire lots may be scrapped unnecessarily. In contrast, with real-time APC, problems with a tool are detected as soon as they occur. For example, if a wafer etched at an incorrect rate, the tool will be flagged immediately and production personnel notified of the problem.

Figure 1: Results of a case study that compared monetary losses associated with various process control strategies. The schemes are presented in terms of their mean times to detection (MTTD) of an etch tool excursion, from lowest MTTD to highest.

This article discusses how periodic etch-equipment qualification procedures were replaced by real-time, on-product monitoring at the 200-mm Infineon Technologies fab in Richmond, VA. The traditional qualification procedure required etching of blanket (or patterned) test wafers of a known thickness for a preset time, followed by a postetch thickness measurement. The process etch rate was then calculated from the film-thickness and etch-time data using SPC software.² With the new technique, etch-time data are collected for certain process steps by an equipment integration (EI) software package that communicates directly with the tool, records readings from in-line metrology devices, and forwards data to the SPC software.³ The etch-time data are then combined with previously gathered pre-etch film-thickness measurements to determine the on-product etch rate. Figure 2 shows this process in a flowchart format. It enables the fab to control an etch tool's performance without a costly break in production for the purpose of running a test wafer. It also provides a continuous snapshot of chamber conditions, which ensures instant detection of out-of-control processes.

Figure 2: Flowchart of the on-product process monitoring procedure that has replaced traditional periodic equipment qualification.

Project Background

In most current etch tools, process recipes are broken into a series of stabilization and etch steps. Typically, an endpoint algorithm is used to determine the completion time of one or more of the etch steps. For these steps, an underlying layer exists that allows the use of an optical endpoint. The etch step ends as soon as the upper layer is removed (see Figure 3). As part of the APC implementation project, the Richmond fab is monitoring etch time for these automatic endpoint steps. Data on etch time as well as film thickness are gathered automatically using EI and SPC software.

Figure 3: Simplified schematic of an etch-rate application. The process consists of two substeps, the first of which uses optical endpoint detection.

Figure 3 illustrates a sample etch-rate application, omitting film types and the complex patterns that the top films typically blanket. In the example shown, the top film is etched in two substeps, the first of which uses optical endpoint detection. This etch removes the unpatterned, blanket portion of the film that is above the recesses of the second film layer. The result is an unetched underlying layer, with the top film still filling gaps between the patterned features. In the next substep, a selective chemistry and timed etch are used to remove a controlled amount of the top film from the recesses. Using the pre-etch measurement of film thickness in an unpatterned area of the wafer and the etch time for the first substep, the on-product etch rate can be estimated.

Project Implementation

Etch-Time Data Collection. To implement on-product APC, the etch equipment to be monitored was configured to report detailed lot-level information to the fab's EI software. Every lot processed on a tool generates a file containing lot ID and recipe names along with the wafer-level parameters that are recorded by the tool. These parameters include wafer number, endpoint time for each optical endpoint step in the recipe, and the specific chamber of the etch tool where each wafer is processed. Thus, endpoint information is available in the EI software's storage files for every wafer. To estimate etch rates, the lot ID and wafer numbers are used to search through SPC software databases and locate the pre-etch film-thickness measurements for those wafers. No additional process time or throughput change is required to obtain the detailed information on etch times, so there is no impact on manufacturing productivity.

One of the key components of the implementation project was to ensure the accurate collection of etch-step times by the EI software, which was revised to meet that goal. To verify the software's capabilities, endpoint etch-time data collected by the fab's computer-integrated manufacturing (CIM) software over a range of dates was compared with an identical set of data obtained directly from a typical etch tool. The results of this comparison can be seen in Figure 4. A steady offset of 0.10 seconds was found at every data point in the 95-wafer data sets. Similar results were also observed when comparisons were performed using data from other plasma etchers.

Figure 4: Comparison of etch times collected using the fab's CIM software and those collected directly from a typical etch tool.

Pre-Etch Film-Thickness Measurement. Before the project was started, film-thickness measurements were taken to control individual film deposition processes. With on-product monitoring, such pre-etch measurements are still used for SPC; however, they also are used for etch-rate calculation. In order to utilize the measurements in this new application, a special pointer variable was implemented in the fab's manufacturing execution system. This variable records the location of the measurement data and stays untouched throughout the etch process. After the etch is complete, the SPC software collects the etch times and uses the pointer variable to access the correct set of film-thickness measurements, which, divided by etch time, will yield the process's etch rate. The ability to use the same measurements for two purposes means that no additional time is required for metrology.

In general, use of a sampling pattern is sufficient to obtain film-thickness data, although the pattern may need to be customized for tools that have multiple chambers attached to a central buffer chamber. For example, on two-chamber tools, a sampling pattern of wafers 3 and 22 may work well. On four-chamber tools, the pattern may need to be adjusted to 3, 14, 16, and 22 to ensure that all etch chambers have a good chance of processing a previously measured wafer.

One goal of this project was to capture data for all chambers on a tool during the course of one lot. Therefore, the sampling process had to be planned carefully, since for a specific chamber's etch rate to be calculated based on on-product etch-time data, a premeasured wafer must have been etched there. The optimized sampling routine provides the greatest probability of each measured wafer being etched in a different chamber. Many tools sequence wafers differently; therefore the metrology sampling plan had to be tool-specific.

Software Configuration. A critical part of the implementation project was configuring the SPC software package to collect, manipulate, and analyze the two main types of data. The software was configured to gather the step-time data after each lot finishes on the etch tool, access the correct film-thickness information, and then calculate the etch rate. The calculated rate is then added to a control chart. If the etch rate is determined to be out of control, the lot is put on hold for evaluation and the problem is investigated.

Etch-Rate Comparison Experiments. To determine whether on-product monitoring had the capabilities needed to replace traditional equipment qualification, on-product etch-rate data were compared with test wafer etch-rate data for a typical etch tool over a 3-month period. The results are presented in Figure 5. In this plot, each X represents a single site on a test wafer. One blanket test wafer was run every 3 days, nine sites were measured across each wafer, and their average thickness was calculated. A trend line was fit to the averages to enable visual comparison. The frequency of the on-product data points was much higher than that for the test wafers, because each point in the set represents an individual wafer that was processed on the high-volume tool. The sampling pattern used for these measurements was similar to that used for the test wafers: nine sites were measured and then averaged for each wafer. The individual wafer etch rates were then averaged by date and a trend line fitted to the date averages. When the normalized etch rates for the two different techniques were compared, the correlation coefficient was 0.66.

Figure 5: Comparison of etch rate (ER) data obtained from unpatterned test wafers used in the traditional equipment qualification procedure and from on-product monitoring.

Both data sets showed a steep drop in the etch rates on the July day when a clean was performed, a known characteristic of the etch process. However, for the test-wafer data set, across-wafer etch-rate nonuniformity increased steadily with time until the clean was performed. In the figure, large variations in test-wafer etch rates can be seen through June and a few days into July, followed by more-uniform across-wafer etch rates. In contrast, for the on-product data set, variation remained similar over time, which is more characteristic of the process tool used for this experiment.

There also was a large degree of spread in the test-wafer etch-rate data. Measurements for each date fell into a bimodal distribution, with center points on the wafer having a lower, more-uniform etch rate than points near the edge of the wafer. This effect, which is dependent on the process tool's wet-clean cycle, can be observed most easily in the data collected during the month of June. The on-product etch rates were more tightly distributed because test and product wafers react differently when exposed to the same chamber conditions. Etch chambers and processes are optimized for product wafers; therefore, the amounts of reactant gases present in the chamber during an etch are set based on the pattern densities and film thickness on the wafer. Because of the absence of a pattern, there are many more reactive sites on test wafers, so when they are exposed to the preset gases the resulting etch is comparatively nonuniform.

When upper and lower control limits were calculated based on the on-product etch-rate data, it was found that no OOC points had been detected during the 3-month period, but the etch rate had varied widely from month to month. Typical SPC methods assume a data set is normally distributed, and in this case, it was not. If control limits are calculated using this false assumption, SPC can still be used, but the limits would be so wide that much process variation would go undetected. However, when on-product etch-rate monitoring is performed, moving SPC limits can be used for chamber control. Such limits are based on the normal change in etch rate exhibited by a weighted collection of previous data points and are recalculated every time a new wafer is run. For example, the etch tool and process used for the experiment shown in Figure 5 has historically had an etch rate that systematically increases and decreases in a sawtooth pattern. The rate increases steadily until a clean is performed on the chamber, which causes it to decrease suddenly. The rate then begins to increase again, until the next clean is performed. The use of moving SPC limits would allow operators to flag minor, but still significant, excursions from this pattern instead of waiting for etch rates to go outside of limits that are set wide to account for systematic variations.

Finally, the data in Figure 5 suggest that instead of performing an etch-rate analysis on every lot, taking a multivariate approach may be more appropriate. Currently, etch time is the only on-product variable measured but it is used in combination with film-thickness data to get an etch rate. In a multivariate approach, instead of estimating and charting etch rate, control charts could be constructed using only endpoint times. If an endpoint was found to be out of control, then film-thickness data could be used to determine whether an etch problem or a film deposition problem had occurred. With this technique, the functions required of the SPC software could be minimized, while still maintaining an excellent level of process control.

When the same type of etch-rate data comparison as that illustrated in Figure 5 was performed for a different process using the same collection and averaging methods, the results (shown in Figure 6) revealed there again was excellent agreement between the on-product monitoring data and the test-wafer data, with a correlation coefficient of –0.84. It is proposed that this negative correlation occurred because the etch reactors were not optimized for the blanket test wafers. It may also be related to this particular etch process. However, the significant agreement indicates that both process control methods detected the same trend.

Figure 6: Comparison of etch-rate data obtained from test wafers and from on-product monitoring. The latter method revealed an out-of-control condition that was not detected using the traditional equipment qualification procedure.

Figure 6 also reveals an important advantage of using on-product monitoring. An unusually high etch rate was calculated using endpoint data on November 5. Upon further investigation, it was discovered that the tool had experienced a pressure fault during this etch. The test wafer for that date, which was run several hours before the fault happened, did not detect any unusual conditions. Using on-product monitoring, the fault was highlighted for the equipment engineer, corrective action could take place, and the possibility of unknowingly exposing additional product to the fault was eliminated. Because it is able to gather more data in a given time period, the on-product technique provides a more accurate estimate of chamber conditions than periodic chamber qualification.

Conclusion

The project to replace periodic etch-tool qualification with on-product monitoring consisted of several components. The fab confirmed the accuracy of the etch-time data collection methods being adopted, planned the sampling pattern in order to give the best chances of capturing the etch rate of each tool chamber during the course of one lot, and configured the SPC software to collect etch-time and film-measurement data and use them to compute etch rates on product wafers. Comparisons of on-product etch rates with test-wafer etch rates demonstrated that the two different methods provide statistically similar results. A correlation coefficient of 0.67 was calculated for one process, and a correlation coefficient of –0.84 was observed for another.

The on-product technique offers many advantages over traditional tool qualification. Data are collected from many more wafers, resulting in better estimates of chamber conditions. Tool availability is increased, since test wafer runs do not need to be performed. Because OOC events are detected much sooner than by traditional means, wafer scrap is minimized. On-product monitoring also has the potential to use advanced moving-limit SPC methods and multivariate analysis.

Acknowledgment

This article is based on a paper presented at the 13th annual IEEE/SEMI Advanced Semiconductor Manufacturing Conference, Boston, April 30–May 2, 2002.

---

References

1. J Schmitz, "Why Would We Accept Less Than 100% Yield? How to Maximize the Amount of Good Dies out of a Given Wafer Fab?" (Keynote presented at the 12th annual Advanced Semiconductor Manufacturing Conference, Munich, Germany, April 2001).

2. Y Karzhavin, "Improving the Etch Process as Part of an Overall Plan to Increase Fab Productivity," MICRO 19, no. 5 (2001): 33–39.

3. T Urenda, "Generic Host Interface Solution" (paper presented at the AEC/APC Symposium XI, Vail, CO, September 1999).

---

Neal Lafferty has worked on lithium niobate static MEMS, optical CD, and, most recently, advanced process control during internships at the Naval Research Lab in Washington, DC, and Infineon Technologies. The author of papers for SPIE, ASMC, and industry trade journals, he received a BS in microelectronic engineering from the Rochester Institute of Technology in Rochester, NY, and is pursuing a graduate degree in materials science and engineering at the same institute. (Lafferty can be reached at nxl7930@rit.edu.)

Bennie Fiol is a member of the team dealing with run-to-run control and a datalog SAS programmer at Infineon Technologies in Richmond, VA. Previously, she spent 10 years as an SAS programmer at the United Network for Organ Sharing. She received a BS in computer science from Mary Washington College in Fredericksburg, VA. (Fiol can be reached at 804/952-7531 or bennie.fiol@infineon.com.)

Paul Jowett is responsible for 170-, 140-, and 110-nm shallow-trench isolation-mask open and poly etches at Infineon Technologies in Richmond, VA. He has more than six years of plasma etch experience. The coauthor of two papers on run-to-run control implementation, he received a BS in materials science and chemistry from Manchester Metropolitan University and an MS in the physics of advanced electronic materials from the University of Bristol, both located in the UK. (Jowett can be reached at 804/952-7208 or paul.jowett@infineon.com.)

Yuri Karzhavin, PhD, joined Infineon Technologies (Richmond, VA) in 1997. As manager of the advanced process control project, he focused on bringing new fully automated advanced control technologies into semiconductor production. Previously, he worked at Motorola's Advanced Product Research and Development Laboratory (Austin, TX), specializing in advanced logic technology development and manufacturing. In 2001, Karzhavin joined the Technologies of Pipeline Transport research institute in Moscow, filling the position of technology and automation director. Karzhavin has authored or coauthored more than 40 publications and manuscripts and holds two patents pending. He received a PhD in plasma physics from Moscow State University and an MBA from Virginia Commonwealth University in Richmond. (Karzhavin can be reached at +7 095 2344505 or karjavine@niittt.comcor.ru.)

Tim Urenda is an advanced process control (APC) engineer at Infineon Technologies in Richmond, VA. Previously, he was an equipment integration engineer at the company and a member of the technical staff at Sandia National Laboratories. He has authored or coauthored papers on fault detection and on APC and factory automation. He received a BS in mechanical engineering from the University of New Mexico in Albuquerque and an MS in mechanical engineering from Purdue University in West Lafayette, IN. (Urenda can be reached at 804/952-7509 or tim.urenda@infineon.com.)