The tremendous impact of APC for plasma etch
10/01/2001
Advanced Process Control
Volker Tegeder, Infineon Technologies SC300, Dresden, Germany
Robert Ronchi, STMicroelectronics, Rousset, France
Sven Mueller, Mathias Hofmann, AMD, Saxony, Dresden, Germany
overview
Collaborative work among several European fabs is proving the value of advanced process control, using SEERS data, for identifying arcing, indicating possible particle excursions, and identifying and verifying possible plasma process improvements. In addition, SEERS analysis has proven to correlate well with electrical test data, making this a valuable early warning tool for keeping a process within specification and improving process yield. As set up, this automated system has a supervisory capability that, with an established fault detection system, can stop a process when there is a high likelihood of a problem.
Although we have been using plasma etch processes for two decades, their subtleties continue to challenge process engineers everywhere. Frequent chamber cleans and associated conditioning processes, process interactions, and arcing all contribute to reducing overall equipment effectiveness. Fortunately, it is especially with the etch process that advanced process control (APC) can make a significant impact on costs, throughput, and yields.
 Figure 1. The backbone of the supervisory control system for SC300's advanced process monitor for plasma etch tools is the Hercules APC NET network. |
The ongoing European Semiconductor Equipment Assessment (SEA) project APC300 is integrating an advanced process monitor with supervisory capability into several different plasma etch tool types (Fig. 1). Participants include, in Germany, Infineon Technologies SC300 in Dresden, Infineon Technologies in Regensburg, AMD in Saxony, and ASI in Berlin, plus STMicroelectronics in Rousset, France. We have completed sensor integration into several different tools and addressed linking necessary logistics data. We have begun to evaluate ways the supervisor can benefit production. After only six months, we have concrete examples of these benefits, including fault detection, more rapid tool recovery and release to production, and improved chamber conditioning. The most important finding is a correlation of the internal plasma parameters with IC electrical test data, enabling us to identify potential yield limitations weeks earlier than before.
Installation
Our primary evaluation site at the SC300 fab consists of several plasma reactors, each equipped with a Hercules plasma sensor and network, and a supervisory system capable of interface with each plasma controller (Fig. 1). The tool set includes tools for three different processes and includes both capacitive and capacitive with additional inductively coupled plasma chambers. The supervisory system enables 100% control over all monitored etch tools up to a total of 12 chambers at one time.
 Figure 2. Monitored collision rate and electron density for a four-step etch process. |
Other IC manufacturers participating in this project have different applications running on single tools of various types. We exchange results between all participants. The investigations at Infineon Regensburg focus on chamber conditioning, dry clean, and process mix issues. AMD Saxony's activities in the SEA project include fault detection and classification as well as reduction in test wafer use. STMicroelectronics is working to improve chamber-conditioning procedures after wet clean.
Generically, the plasma sensor mounts the same way to each plasma tool; this is via a passive sensor head on a tool's chamber. Of course, each specific tool requires a unique interface configuration. Initially, we integrated the passive sensor head and then its control system, which uses self-excited electron plasma resonance spectroscopy (SEERS) to characterize the bulk plasma.
Some reactors use a rotating magnetic B-field to increase electron lifetime in the plasma, enhancing etch rate. The varying magnetic field adds an oscillation to SEERS data, and generates a noisy signal. In these reactors, we had to synchronize the trigger signal of the Hercules system to the rotating field to strip oscillations from SEERS measurements.
The next challenge was to create a link so the supervisor could access the logistical data needed for data analysis and fault detection. These data include lot identification, recipe, and wafer-slot identification. In our first installation, we used an "espionage" technique. By monitoring the tool's SECS communications with a "SECS Spy" made by ABAKUS Software GmbH, Dresden, Germany, we were able to pick out the needed data and correlate them with sensor data within the supervisor. While perhaps not the ideal way to gather data, this method did not require any equipment changes.
In the second installation, we were able to take advantage of LAM's innovative plug-and-play tool-sensor interface. Using an Ethernet-based 10Mbit independent network, the supervisor has access to necessary logistical data.
Theory of operation
The plasma sensor is mounted into a flange to assure that the inner chamber wall is approximately in line with the detector. Briefly described, the plasma sensor measures RF discharge current that is analyzed to determine electron density and collision rate. By using a general nonlinear gas-discharge model, the APC system provides reciprocal and volume-averaged values of electron collision rate and electron density (Fig. 2), and bulk power dissipated in the plasma body.
Like collision rate, electron density is a global plasma parameter. Collision rate is very sensitive to process conditions as shown by
 |
where nstoch = stochastic heating of the sheath at the electrode, Te = temperature of the electrons, p = pressure, pk = partial pressure of gas k, TN = gas temperature (neutrals), and sk = particle size (cross section).
Electron collision rate, describing the interaction of electrons and process gas, is one essential parameter for modeling and controlling plasma processes. Beyond the influence of electron temperature (Te), collision rate also depends on the density of the process gas resulting from gas pressure and temperature (p/kTN), so it is a natural parameter in contrast to pressure defined by a recipe. Collision rate has another important dependence, that is, the effective concentration of different gases and particles generated by the plasma itself.
Collision rate is directly affected by changes in temperature, pressure, and RF power. Since the cross section (sk) is strongly dependent upon the size of gas molecules, collision rate is also affected by the presence of large molecules, such as polymers created as a byproduct of an etch process. This sensitivity can pick up changes in chemistry, chamber conditions, prior processing, and arcing. As we shall show, we found that we can use SEERS data for fault detection and tool start-up testing.
The plasma potential (i.e., the voltage drop of the dark space at the chamber wall) is usually about 20V. The voltage at the sensitive but passive sensor electrode is ~10mV. Thus, the sensor cannot affect the plasma process. Deposition on the sensor does not affect results because the deposited film creates a minor series impedance (<<50W) that can be neglected using this technique.
Process fingerprints
We used the installed sensors and APC system to fingerprint each process. A fingerprint consists of average values of collision rate and electron density for the process. With such data, we could compare the average value to any process run to identify abnormal conditions and to optimize any of our processes, as seen in the examples below:
 Figure 4. A collision rate lower than the system fingerprint indicated arcing at the gas distribution point. (Source: AMD) |
Arcing. Data presented in Fig. 3 show a high collision rate for days 13 and 14 of a test period. We thought this might indicate an arcing problem with this tool, so we shut it down for investigation. Arcing at the chamber wall becomes more probable with increasing polymer thickness and depends on both chamber age and the recipe. Indeed, we found that the chamber was arcing as indicated by arcing traces on parts of it. Large polymer particles, injected into the plasma during the arcing at the chamber wall, led to an increased collision rate.
In another situation, we detected a collision rate lower than the fingerprint, in a new tool at start-up without any polymer on the chamber wall. By identifying this problem quickly with SEERS data (Fig. 4), we were able to shorten start-up time considerably.
System warm-up. When we compared two plasma etch recipes with very different gas flows and dissipated power, the sensor showed similar warm-up behavior before the etch processes stabilized. This gave us justification to set a specific number of warm-up wafer runs prior to restarting production. We may see a yield improvement as a result of this change if insufficient warm-up has any impact on product wafers.
Chamber conditioning. Our analysis of chamber conditioning after a wet clean gave us a mechanism to improve the process directly. One process we looked at required a particle count <50 before production lots could begin. This typically took 50 conditioning wafers. To begin our investigation of this situation, we correlated collision rate with particle counts and found that when collision rates are low there are fewer particles on the wafer.
With this correlation in hand, we were able to quickly assess the benefits of any potential process changes. Our initial process with 50 conditioning wafers used resist-coated wafers followed by wafers coated with polysilicon to precondition the chamber. First, we replaced all these wafers with patterned oxide wafers. This change reduced collision rates and particle counts, enabling us to use only 37 conditioning wafers. However, we felt the process was still far from optimized.
Reverting back to just five resist-coated wafers followed with five blank oxide wafers, we were able to meet our <50-particle limit with just 10 wafers, instead of 37 or 50 (Fig. 5). We also found that after this condition, collision rates for production lots were significantly lower than with our initial process, indicating a further conditioning of the chamber with product wafers.
Electrical test. One of the most promising revelations of our investigation was our ability to correlate SEERS data with electrical test data. In particular, gate contact length bias correlates very well with collision rate (Fig. 6). This ability to foresee electrical results 14 days in advance gave us a powerful tool to improve yields. As we understand what gives rise to collision rate changes, we can use the APC system as a tool to control these changes, which should improve our control over gate contact length and increase yield.
In other words, a high collision rate indicates that a method is approaching its process window limits.
Throughout our investigation, we found that as long as the measured plasma parameters remained within a small window, our process parameters stayed within specification. As plasma parameters began to deviate, the process shifted closer to the edges of the process window.
We found that the stronger the shift in plasma parameters, the stronger the shift in process parameters. We concluded, then, that our APC system could be used as an early warning system, alerting us to changes that would eventually throw the process out of specification.
Conclusion
Although we are only six months into our evaluation cycle, our APC system has already proven itself to be an effective tool for identifying arcing, indicating possible particle excursions, and identifying and verifying possible process improvements. We have demonstrated a test wafer reduction from 50 wafers down to 10 wafers after a clean by using SEERS data.
Using supervisory capability, we will be able to set up decision trees that can stop the process when there is a high likelihood of a problem. Since results of SEERS analysis correlate so well with our electrical test data, it is likely that this tool can help us improve our process yield. We have also seen that we can use sensor data as an early warning system to keep our process in specification.
Acknowledgments
The following authors also contributed to this article: Matthias Scholze, Torsten Prescher, Uta Federbusch, Thomas Bauer, and
Eckhard Marx of Infineon Technologies SC300; Ute Nehring of Infineon Technologies, Dresden; Mark Laqua of the University of Dresden; and Michel Derie, STMicroelectronics, Rousset, France. This work is funded by the European IST program 4-8-3 on the SEA initiative.
Volker Tegeder received his PhD in semiconductor physics from the University of Ulm in Germany. He is responsible for the introduction and evaluation of new methods of process control for enhanced productivity at Infineon Technologies, SC300 GmbH & Co. OHG, Königsbrücker Strasse 180, D-01099 Dresden, Germany; ph 49/351-886-7351, fax 49/351-886-7052, e-mail [email protected].
Robert Ronchi holds a PhD and is equipment strategy and engineering manager at STMicroelectronics in Rousset, France.
Sven Mueller studied electrical engineering at the FH-Lausitz with a focus on microsystems and automation. He is a manufacturing engineer in the etch module at Advanced Micro Device's Fab 30.
Mathias Hofmann studied experimental solid state physics. He is a senior process engineer at AMD, Saxony.