Advertiser and

Buyer's Guide
Buyers Guide

Chip Shots blog

Greatest Hits of 2005
Greatest Hits of 2005

Featured Series
Featured Series

Web Sightings

Media Kit

Comments? Suggestions? Send us your feedback.

Taking Control

Devising an APC strategy for metal sputtering using residual gas analyzers (Last in a series)

Gerald Rampf, Infineon; and Robert McCafferty, Curvaceous Software

A study used RGA technology coupled with tight sensor integration, key-number compression, and multidimensional analysis to eliminate the introduction of contaminants into sputtering tools.

A great deal of manufacturing effort is invested in wafers that reach the metallization stage. At that juncture, the penalty paid for producing scrap—not only in terms of dollars but also dissatisfied customers—can be excruciating. Such lots are very close to the testing stage, and starting additional material to satisfy immediate demand is simply not feasible. However, sputter deposition processes can result in scrap; they are performed in high-vacuum chambers and can be extraordinarily sensitive to contamination (particularly of an organic nature), which can quickly lead to electromigration and potential field failure issues if it goes undetected. The problem is that given the nature of argon-driven metal sputtering using aluminum, aluminum/copper alloys, or titanium/titanium nitride alloys, the process has been largely unmeasurable in a manufacturing environment.

Research at Infineon (Dresden, Germany), however, has armored that Achilles' heel. As detailed in this article, that research involved the novel use of residual gas analyzer (RGA) measurement methods coupled with very close tool/sensor integration, key-number data compression to deal with a small tsunami of raw data resulting from effective sensor integration, and multidimensional techniques. These multidimensional techniques effectively yielded a "process camera" that could smoothly handle the torrent of apparently unrelated but in fact physically coupled numbers resulting from successful sensor integration. This article focuses on the technical work underpinning this project and the surprising benefits to be gained from expending integration, engineering, and analysis effort as well as R&D funds on incisive, process-dependent fault detection.

Why Measure?

With five levels (and more) of metallization commonplace in the global semiconductor industry, where only a handful of highly competitive DRAM manufacturers remain, the game has changed. End-of-line (EOL) sorting has given way to scrutiny of in-process as well as in-line parametric data with advanced process control (APC), which consists of fault detection and classification in tandem with run-to-run control. The new methodology is applied at every juncture where insightful measurement is possible and mistakes are dear, including metal sputtering, which long ago became the method of choice for interconnect deposition at Infineon.

In the metal sputtering process, argon gas is bled into an evacuated chamber and RF plasma is struck. That plasma, as depicted in Figure 1, accelerates Ar+ ions across its sheath potential to strike a powered metal target of desired film composition, physically dislodging (sputtering) and effectively blasting atomic-level chunks of target material throughout the sputtering chamber. Some of this material successfully deposits on grounded target wafers to produce a contiguous, high-quality metal film that, under ideal conditions, completely fills all desired interconnect pathways.

Figure 1: Chamber-level schematic of metal deposition through argon sputtering.

Despite its technical elegance, this method is obviously prone to disturbances, particularly from unwanted contaminants that alter interconnect film properties or otherwise raise electromigration and corrosion concerns. Under the millitorr-level pressure conditions of vacuum sputtering, such failure modes take a variety of forms, including air leaks, which lead to the formation of oxide and nitride constituents in deposited films. Even impurities harbored within the voids of metal target welds constitute a threat. Yet, these contaminant forms cause far fewer problems than the wholesale introduction of organic compounds from wafers with trace amounts of incompletely ashed photoresist films. In addition to generating costly scrap, incompletely ashed films can cause a sputtering system to suffer substantial downtime. The purity of gas supplies and targets is also a perennial concern.

Product failures stemming from these contamination sources can range from obvious and explicable in the photoresist case (presuming all tainted work-in-progress wafers are successfully isolated) to insidious and expensive field reliability failures rooted in trace inorganic contamination. Beyond the dollar amount of dead wafers and distrustful customers, however, process failures go against the grain of known best business practices by forcing manufacturers to maintain high inventories, resulting in excess costs. In addition to having undergone up to 90% of their total processing by the time they reach final metal deposition, wafers at that stage are approaching the testing phase, when replacing ruined work-in-progress wafers with newly started material would make it impossible for manufacturers to meet customer commitments. Despite these manifest dangers, contamination levels in metal sputtering processes typically are not tracked. In no small measure, that deficiency is a result of the inherent technical difficulties involved in implementing tracking strategies. The consequent leap of faith involved in not investing in advanced process control converts what was a toss of the dice in 200-mm manufacturing into effectively betting the farm in 300-mm manufacturing.

Half the Battle

Sensor Selection. Finding a mechanism that is sensitive, robust against disturbances, cost-effective, reliable, and readily maintainable to detect the types of contamination that can plague sputter processing is at least half the battle. Making it work in manufacturing is the other. For Ar-driven metal sputtering processes, however, a cornucopia of options does not exist. The conventional method of tracking tool parameters (power, pressure, flow, etc.) through the SECS port does not offer insight into chamber contamination levels. While optical methods are possible, since plasma emissions provide strong light and a viable signal source at discrete frequencies corresponding to plasma components, they lack sensitivity at the trace level, where sputtering contaminants become a problem.

The search for adequate sensor methods led to the use of RGAs, which were well known at Infineon for performing helium leak detection in physical vapor deposition tools. Operating on the principle of ionizing reaction-chamber gases and then separating the results by mass-to-charge ratio, RGAs can detect and distinguish between small amounts of impurities, thereby providing excellent sensitivity and resolution for process chamber applications. As diagrammed in Figure 2, ionization is accomplished by means of electron bombardment, with ion separation and counting occurring in quadrupole rods followed by a detector. RGAs come in three basic configurations:

1. Open ion source (OIS) sensors. Reasonably inexpensive, these garden-variety RGAs are well known to Infineon's maintenance staff, but have a cutoff point of approximately 10–5 Torr, the maximum measurement pressure permissible without causing instrument damage. Sputtering process pressures, however, are in the millitorr range.

2. Closed ion source (CIS) sensors. Basically a modified version of OIS devices, these RGAs introduce sampled gas through a small orifice and engage a turbomolecular pump to maintain low operating pressure within the quadrupole rods. That mechanism yields sensitivity up to atmospheric pressure but results in a much larger, more shock-sensitive, pricier assembly.

3. Extended pressure range sensors. Featuring a minimized mass separator (but still based on quadrupole rods), these small and handy RGAs can operate at pressures of up to 10 mTorr. However, they have limited lifetimes because of analyzed contamination and are fairly fragile, making them less than ideal in a manufacturing setting.

Figure 2: Schematic of residual gas analyzer operation.

Unfortunately, in situ measurement evaluation revealed that at process pressures, CIS sensors experienced signal swamping from high Ar flows and thus offered limited sensitivity to the contaminants being sought. Nevertheless, that setback turned out to be the mother of considerable invention.

Realizing that contaminants would still be present in the sputtering chamber long after the process gas (argon, in this case) had been switched off but before pumpout for the next wafer was complete, the engineers attempted to perform measurements between wafer runs. That decision led them back to OIS sensors, which are more familiar and maintenance-friendly than CIS sensors and, fortuitously, were already available on Infineon's sputtering systems. As shown in Figure 3, the sensitivity of the OIS sensors was excellent on all atomic mass channels analyzed. The same OIS instruments, as indicated in Figure 4, could even be used at normal process pressures to detect the introduction of photoresist into the sputtering system degas chamber (albeit at long process times, since pyrolysis is necessary). With a little imagination, the engineers found a workhorse sensor for advanced process control but now faced the problem of integrating that sensor with the sputtering equipment set.

Figure 3: The sensitivity of the OIS sensors was excellent on all atomic mass channels analyzed in between wafer runs.
Figure 4: The use of OIS sensors at normal process pressures to detect the introduction of photoresist in the sputtering system degas chamber.

Sensor Integration. The RGA integration problem broke down into two components:

1. Hardware integration, which was concerned with reliably getting relevant signal data from the sputtering system without compromising either tool functionality or RGA longevity.

2. Data integration, which was necessary for associating raw RGA data gathered between wafers for each atomic mass channel with relevant process data and, after key-number compression, with lot ID and wafer test results.

Since the RGA is a passive device mounted downstream of the process chamber and is inactive during sputtering, it effectively cannot compromise processing integrity, barring a major vacuum breach through the RGA itself. Hence, the hardware integration problem reduced itself to creating a special interface to ensure that the OIS sensor induction port was shuttered off when either process gas was flowing or chamber pressure was rising. For reliability purposes, hardware was the best candidate for creating that interface, since all software packages (particularly Windows-based varieties) are prone to unexpected behavior in a manufacturing environment. The data integration problem had two components: (1) process-data acquisition and synchronization, and (2) compression followed by linking process data to lot-level test results.

Process-data acquisition and synchronization, executed at the tool level, required logistical information to correctly associate wafer-level postsputtering RGA measurements to process tool parameters, such as power, pressure, and flow. Since the sputtering system's SECS port was connected to the fab host, the acquisition of process data necessitated the construction of a passthrough board (diagrammed in Figure 5), which served to transmit SECS messages back and forth. The passthrough board did not impede the flow of messages between the host computer and the sputtering tool. To the contrary, it enabled the computer to synchronize and store tool and RGA information as well as lot, wafer, and recipe ID.

Figure 5: Hardware arrangement to accomplish tool-level data integration.

Once the hardware-driven sensor and data integration scenario depicted in Figure 6 had been established, the final step in the process control strategy involved compressing the data and linking them to EOL results.

Figure 6: Overall RGA sensor integration at the tool level.

Surviving the Flood

Successfully integrating the sputtering systems at Infineon with RGAs greatly compounded the magnitude of linking tool data and EOL test results. Far from basking in their successes, the architects of effective RGA integration were now confronted daily with more than a thousand files of raw, real-time data per tool. Manual analysis was utterly untenable, while the basic objectives remained unattainable—detecting failures and determining for each wafer entering each chamber whether the sputtering system was processing correctly.

Since only a go/no-go decision for chamber viability was required, rather than a detailed diagnosis of tool problems (which could always come later if issues were found), the premier goal was to compress time-variable RGA data to a smaller set of measurements effectively capsulizing chamber condition. Achieving that goal required the application of standard key-number compression techniques, where time-variable molecular-weight signals from the RGA are decomposed into minimum, maximum, mean, and standard deviation values reflecting behavior over the entire measurement period. This procedure condensed hundreds of potential measurement points per wafer into only four key numbers for eight RGA mass channels, or 32 values per wafer.

Key-number compression helped immensely to reduce the sea of data files to something on the order of a lake—still a very large lake that included large bodies of high-dimensionality EOL data for each sputtering chamber. The standard approach in such cases is to perform compression-by-physics to isolate a subset of RGA key numbers that can then be mapped forward to EOL results and backward to tool and process parameters. Compression by physics provides a small set of information-rich parameters that, given reasonably successful modeling results, can be monitored in proxy for a large number of real-time values as well as EOL results. This technique has worked well for etch processes, which have been the subject of an extensive, plasma-centered study over the past decade.1 Unfortunately, the technique has not yet worked as well for sputtering processes as it has for etch processes. Hence, in addition to 32 RGA key numbers, the Infineon engineers wrestled with tool and process data that had to be integrated with EOL test data, mined for process information, and then worked into a control system.

Final data integration—the lot-level pairing of compressed RGA measurements and EOL test numbers (or yields against electrical specifications)—enables not only process optimization but also inclusion of EOL results in tool-level control. This procedure began with RGA trend files (resident in the RGA computer and in the APC trend database) and product data files. Both of these sources contained time-stamped, wafer-level data that could readily be compressed into wafer-level key numbers and then further compressed over a time range into lot-level data using logistical information derived via the SECS passthrough board. RGA and process tool data were merged over common time ranges, while lot IDs provided a key for matching those data with product-level test data from EOL sources. The result of this process was a complete, albeit a rather broad, multiparameter picture of sputtering system behavior that included tool data, RGA data, and EOL test results. These data were compressed into a usable format and density at both the wafer level (suitable for tool go/no-go decisions) and the lot level (useful for process optimization and control). The data-flood problem was solved, but the question remained: How could such large bodies of high-dimensionality data be used to make incisive processing decisions?

Making Sense of It All

Problem. The compressed RGA data alone, which would provide a window into the physics of what was transpiring in the process chamber, required that the engineers deal with a full 32 channels of information. No physical basis existed for distilling that information into a single signal or set of signals, capturing the true pulse of what was occurring at the wafer surface. In order to make informed manufacturing decisions, such as whether a process chamber was viable or had just been fouled by photoresist, the engineers had to deal with information diffused across a large number of variables. Although many of these variables may have been insignificant in themselves, they were in fact crucial because they interacted with other variables. That trait defeats truncation approaches (which ignore apparently uninfluential variables) and renders unworkable conventional methods involving exploration or testing of all data with line, scatter, or statistical process control charts. To deal with the wealth of high-dimensionality information that sensor integration brought to sputtering operations, new technology tailored to deal with multidimensional data was needed.

Solution. To solve that technology problem, the Infineon engineers turned to the notion of parallel coordinates, which has gained rapid acceptance outside the semiconductor industry.2,3 As diagrammed in Figure 7, the parallel coordinate method operates by converting N-dimensional information (such as the 32 channels of sputtering data) into a 2-D representation via a coordinate transformation. In contrast to x, y, and z coordinate axes laid out orthogonally to one another—a method based on a conventional understanding of physical space common to all humans and used by engineers to represent data—the coordinate axes in the novel coordinate technology are parallel to one another. Not being restricted by orthogonality, parallel coordinate technology can use as many variable axes as necessary to describe a problem (e.g., one for each of the 32 channels of RGA data). Values for individual variables are plotted on their respective axes and joined by line segments to form a contiguous, polygonal line. Consequently, a single 16-dimensional observation, isolated from a set of more than 400 observations containing semiconductor product engineering data, can be plotted as in Figure 8. Moreover, when a large number of observations are plotted, such as those that are encountered when dealing with wafer- or lot-level data, patterns form that are detected by the human eye—a superb pattern recognizer. By applying a clustering algorithm to each set of RGA key numbers in turn, the simple mechanism of parallel coordinate analysis unveiled the type of wafer-level sputtering data charted in Figure 9.

Test Results

Many facts, such as the tool fault conditions evident in Figure 9, were unearthed in the nearly 20,000 records of wafer-level RGA data represented in parallel coordinates in Figure 10. The focus of the Infineon study, however, was on developing a comprehensive fault-detection mechanism to automatically assess chamber viability for incoming wafers and react accordingly. To do that, the engineers culled from the full body of data reflecting the sputtering system's historic performance all data indicating adverse conditions or abnormalities, thereby separating out data representing healthy wafer performance.

Visual queries were used to isolate, investigate, and resolve anomalies that became apparent when the data were presented in parallel coordinates, as in Figure 9. Observations not representing best operating performance were excised, thus yielding only those data comprising the best operating zone for the sputtering system, as shown in Figure 11. Every black line (i.e., observation) in that 2-D portrayal of multidimensional sputtering results represented valid tool operation. (Additional data existed in the form of product test results.4,5) Taken as a group, that black mass of data therefore defined the multidimensional shape of desirable tool operation. Two types of information, drawn from the historical sum of all valid operations, were encoded there:

1. Extreme limits, where uppermost or lowermost black lines crossed the vertex of the variable axes. In previous runs, no RGA key-number measurements exceeded those values and still resulted in good tool operation.

2. The interrelationship between RGA key-number measurements for valid operations, depicted by an outline of black lines between adjacent variable axes.

Information on interrelationships is particularly important, since it allows engineers to determine when any variable, which may be well within its historically valid limits, is no longer in sync with all other variables, causing faulty tool performance. That feature of parallel coordinate analysis adds enormous sensitivity to fault-detection and run-to-run control schemes because it enables the derivation of working limits (the green lines in Figure 12) within extreme limits (the red lines) for any configuration of current measurement values (the blue points). This method of deriving working limits allows much more processing latitude than does the use of fixed limits, which, by definition, cannot consider the ramifications of current data (hence, tool state) for all possible variable interactions and must be extremely tight to generate desirable results under all conditions.

The "process camera" for the sputtering system depicted in Figure 12 (which is actually a combined fault detection and run-to-run control vehicle) indicates that the M29 and M37 variables exceeded working limits. Taken together, these variables characterized an underlying event that drove the sputtering system to an alarm condition. Treating each variable as a letter, engineers could mechanize future responses by spelling out the names of such alarm conditions, cataloging them, and recording how they were resolved. If a condition is associated with or remedied by process variable manipulation (e.g., off-target incoming wafer properties such as film thickness, necessitating process adjustment), the minimum control move to rectify the root problem can be directly, and in fact mechanistically, deduced.


Infineon has invested substantial resources in advanced process control research in an effort to prevent processing errors. That effort took the form of RGA measurement technology coupled with tight sensor integration, key-number compression, and multidimensional analysis to eliminate the introduction of contaminants, particularly photoresist, into sputtering equipment. The findings presented in this article may be applied to other types of equipment, most notably furnaces. In particular, the research discussed here has proven invaluable to 300-mm manufacturing and may lead to the implementation of control strategies for 300-mm tool sets and processes. Benefits derived have not only been proportional to wafer cost at metal deposition and start rate, but also have facilitated the effectiveness of less-senior operating personnel in driving potentially error-prone processing.


1. U Nehring, A Steinbach, and R McCafferty, "Linking Process Parameters to Tool Parameters and End-of-Line Results," MICRO 20, no. 5 (2002): 23–30.

2. R Brooks, R Thorpe, and J Wilson, "Geometric Process Control for Improved Alarm Management" (paper presented at AIDIC, Florence, Italy, May 2001).

3. A Inselberg and B Dimsdale, "Parallel Coordinates—A Tool for Visualising Multivariate Relations," in Human-Machine Interactive Systems, ed. A Klinger (New York: Plenum Publishing, 1991).

4. G Rampf and R McCafferty, "Sputter Chamber Instrumentation, Sensor Integration and Data Acquisition for Tool and Process Fault Detection with Linkage to EOL Results" (poster presented at Sematech AEC/APC Conference Europe III, Dresden, Germany, April 10–12, 2002).

5. R Brooks and R McCafferty, "The Picture of Normality" (poster presented at Sematech AEC/APC Conference Europe III, Dresden, Germany, April 10–12, 2002).

Gerald Rampf is a project coordinator for sensor integration in the CVD/PVD area at Infineon's 300-mm manufacturing facility in Dresden, Germany. He has been responsible for selecting and implementing nonstandard sensors for manufacturing equipment and for acquiring and synchronizing their output with tool and product test data. Prior to gaining his 300-mm experience, Rampf worked as a system expert for equipment engineering and as a project engineer for continuous improvement projects at Infineon's 200-mm line. He received a degree from the University of Technology in Chemnitz, Germany. (Rampf can be reached at +49 351 8867450 or

Robert McCafferty operates RHM Consulting as the North American agent for Curvaceous Software. He began work in the semiconductor industry at IBM Microelectronics in Burlington, VT, specializing in the development and implementation of adaptive control. He has also consulted for a subsidiary of Bolt, Beranek, and Newman, which subsequently became part of Brooks Automation. He received a BS and MS in mechanical engineering and a masters in computer science from the University of Virginia in Charlottesville. (McCafferty can be reached at 203/270-1626 or

MicroHome | Search | Current Issue | MicroArchives
Buyers Guide | Media Kit

Questions/comments about MICRO Magazine? E-mail us at

© 2007 Tom Cheyney
All rights reserved.