Advanced Process and Equipment Control Enables New Device Nodes

Advanced process and equipment control (APC/AEC) has experienced a rapid evolution. As the industry continues progressing toward increasingly complex and unforgiving processes of shorter durations and higher criticality, some form of APC/AEC must be included for yield success, which is why all major semiconductor OEMs offer tools with a form of it. Fabs have changed their infrastructure to accommodate control schemes that require information from multiple sources and at different instants in the various processes to feed forward/feed back data to various tools, and maintain control of the diverse steps.

The APC/AEC path hasn't been smooth. "An ongoing issue is a lack of standards," said James Dougan, principal staff engineer at Freescale Semiconductor (Austin, Texas). "It's an issue at 200 mm; however, at 300 mm, there's greater awareness and adoption of some of the new Sematech standards. Leveraging the data into a meaningful form to do tool-level control is where the biggest challenges lie."

Dean Dawson, senior director of marketing, automated AFM systems, at Veeco Instruments (Santa Barbara, Calif.) agrees. "More control is needed in many areas: CD control, etch depth control and CMP tools."

With the adoption of APC/AEC, semiconductor manufacturing equipment must be able to provide data on demand to enable effective data management and process control. Factory-keyed equipment, such as the Spartan Sorter, offers a critical automation capability to speed tool deployment and provide IC manufacturers with access to the tool data required by advanced fab automation architectures. (Source: Asyst)

The complexity of processes and tools demands an early dialogue between fabs — as tool users — and OEMs. As Dougan put it, "Tool-level data is fundamental to APC/AEC. For example, many tools have sensor data that hasn't always been available to the process engineer. A better understanding of the slight perturbations in sensor behavior and response would help understand and reduce process variability. Many problems arise from not getting data from those sensors in a timely manner."

Another aspect is that the tool must return sensor data at an adequate rate. "We've worked on voltage waveform control for copper plate," Dougan said. "Historically, some information comes off through standard communication ports with an internal refresh rate of about 1 Hz. We found changes in the way that tools perform by looking at samples at the kilohertz frequency level. A 1 Hz sampling rate may be okay in most instances; however, when we consider the subtleties of what can happen within a 1 sec time period in nucleation processes, sampling frequency must be significantly increased to detect slight changes and respond within a short time period." This means multiple Hertz sampling ratios, and the spectrum must go from seconds to thousandths of seconds to output a signal of significance to the process and to allow compensation for variations (Fig. 1 ).

1. Process perturbations must be detected and controlled practically immediately to reduce risk to the product. Driving product yield and performance based on advanced process controls (APCs) at the point of processing is of prime importance. This can only be achieved by accessing and analyzing a higher region of tool-level data. (Source: Freescale Semiconductor)

There is a gap in mining tool-level data. Fab process tools collectively output terabytes of data — some at second or millisecond rates. Processing such data should occur as near the point of occurrence as possible and be effectively summarized. This should then enable other levels of data mining, making it possible to measure, compare and contrast to get information on yield, inline measurement, parametric data or other factors. However, raw data should be made available for true root-cause determination.

"A major gap is that although there are many internal tool control schemes in their process tools, OEMs don't always provide a means of accessing that information," Dougan said. "We need access to this higher level of tool control and data. Without it, meeting processing needs becomes more complicated. When suppliers develop applications for the process tools, we get a recommended process, but not a control scheme around which we can control that process."

Alternative metrology systems can be used in the fab to determine whether a process is adequate. Standalone metrology tools are the verification means that the process tool and its feedback need for stability. However, this requires correlating the two level-control schemes and obtaining metrology data to determine if there is process drift. This requires more raw process tool data, in conjunction with traditional metrology data and feedback.

Pendleton views standards as a way to increase data availability for APC/AEC. "We've implemented the new E-Manufacturing SEMI Equipment Data Acquisition standards [Interface A] to provide data directly to applications for real-time decisions," he said. "We're also coordinating efforts with the Process Control System Taskforce, a SEMI working group. As part of our work with the taskforce, we've developed proof-of-concept demonstrations that show how data may be provided from Interface A to APC-type applications following SEMI Standard E133." Pendleton also sees a need to provide additional data to IC manufacturers to enable decisions, such as run-to-run control and fault detection classification — all necessary for effective APC/AEC programs.

There is a longer-term data mining issue. It is somewhat like genealogy — one must go back and understand every step that a particular lot has taken from fabrication through assembly and test. This requires data traceability for trend detection and other more long-term issues. "Data alone is nothing," Pendleton said. "It must be classified to enable decisions that improve the process." A difficulty is that everyone looks at it differently, depending on where the root of the problem is perceived to be. "What it comes down to, especially at <45 nm, is process variability," Pendleton said. "If it isn't controlled, yields vanish."

As Veeco's Dawson pointed out, since the 65 nm node, optical critical dimension (OCD) has been increasingly used inline to target average CD. It helps etchers achieve correct values. "OCD's issue is that, although it offers benefits, it has some bias issues and requires monitoring," he said. "AFM can be used to control or verify the models used for OCD. Scatterometry can also control OCD's bias effects."

Atomic force microscopy (AFM) is used to control etch depth for feed-back and feed-forward APC. An example is backend structures in Metal 1 low-k. Other technologies may struggle because of charging issues or because the OCD uses a model that may not hold up well. AFM is primarily a direct measurement for depth. While OCD determines pass/fail, AFM provides the reason why there was failure. It can also give a range — minimum and maximum depths — whereas other techniques provide averages.

Dawson pointed out that shrinking technology nodes require a tool mix for control. "Scatterometry has a place, and CD-SEM has been a CD control mainstay, but now 3-D structures and LER, undercut, etc., are becoming critical," he said. "There's an increase in gate metrology requirements. SOI is common, as are stressed films. As chipmakers come up with tricks to get devices to execute faster, such as implants and stressing the device, controlling stress becomes important, and there are numerous ways to measure this."

Scatterometry has its place, but it is complicated, and there is some pull coming for development into an inline APC to measure these stressful films at the gate. Profile, once less important than sidewall angle, must now be monitored on these complex structures. It becomes a 2-D/3-D measurement that is tough to do from the top down. It is almost necessary to look at the side of the device (Fig. 2 ).

2. As features become smaller and more complex, the need to measure factors such as linewidth roughness (LWR) and line edge roughness (LER) increases. 3-D AFM provides the critical capability for 45 nm advanced APC monitoring. (Source: Veeco Instruments)

"Device shape is important," said Wayne McMillan, director of marketing for optical CD at KLA-Tencor (San Jose). "Optical CD is used in production today, not only for CD control, but also for device shape. APC is used at the 90 and 65 nm nodes. There are two APC loops in production today. The first is the gate, CD control. Device makers measure the gate with optical CD after litho and usually do 100% wafer sampling. After litho, those results are fed forward to the etcher, and on it they have a trim curve. For advanced devices, they reduce the transistor's CD using an etch trim, which thins the resist to reduce the CD to obtain smaller gates."

A second optical CD APC loop measures copper volume in the chemical mechanical planarization (CMP) process; traditionally, copper thickness after polishing would be measured by a metal thickness tool. The problem is that metal thickness alone provides no data about the interconnect's overall electrical performance; thus, the correlation between metal thickness and electrical performance is not significant. With OCD, not only is thickness measured, but the CD also is used to calculate the copper's cross-sectional volume in the line. This correlates well with the copper lines' sheet resistance, with cross-section TEMs, and can predict electrical performance. Device makers polish the wafer to what they think the nominal condition is and, after CMP, measure every wafer and check the cross-section volume. If the volume is not on target, an APC loop adjusts the polish rate to improve the copper lines' overall electrical performance (Fig. 3 ).

3. The measurement of 3-D structures, LER and undercut is necessary to maintain processes within their windows and obtain desired yields. The application of OCD to measure the volume of copper lines is a useful technique in this respect. (Source: KLA-Tencor)

Inevitably, there will be many more APC loops at 45 nm. This will result in more APC control, not only from litho to etch, but also from litho to hard mask etch. There will probably be more APC control loops in film thickness to CD, as well as for the spacer formation on the gate, which is critical because it determines where implants go and what the transistor looks like. Here, it is not only spacer width that is important, but also the height and pull down, which determines how much silicide will be present after the gate's silicidation.

There is a balance between how well a tool provides process uniformity, how much metrology is needed for this, and how much APC is needed on that metrology. These are somewhat opposing forces. Process uniformity continually improves, so there is a metrology trade-off. Where processes do not improve, integrated metrology is necessary. Many chipmakers have asked about integrating OCD, but this has not been fully implemented. Process uniformity is improving at such a rate that often standalone metrology with APC seems sufficient to hit process windows.

Life is difficult for the metrology provider. The various materials being introduced, sometimes without full characterization, pose problems. Shrink needs can usually be met by increasing sensitivity. Materials, however, can be unique depending on the chipmaker. Also, in the back end, there are now more layers that are at tight pitches. Metal 1 was once critical, with the other layers being more relaxed; however, Metal 1 through Metal 6 or 7 layers are now becoming tightly controlled. Some chipmakers are monitoring Metal 1 to 7 with OCD for litho, etch and CMP.

Noam Shintel, corporate marketing manager for Nova Measuring Instruments (Rehovoth, Israel), considers the move to 45 and 32 nm as challenging with the introduction of new manufacturing paradigms, such as double patterning and immersion lithography, and new materials, such as metal gates, nickel silicide (NiSi), high- and low-k dielectrics — all of which require tighter process control on high-throughput and highly automated process tools. "In this environment, APC proliferation into new areas like lithography is inevitable," he said. "We'll see APC on parameters like fine shape profile — SWA of a vertical profile, foot and notch."

APC metrology must provide significantly higher throughput to allow more wafers to be measured and more sampling on each wafer to cover across-wafer variations. Tighter process windows pose precision requirements on metrology tools an order of magnitude tighter than the process step error budget. However, precision on a single tool is insufficient; fleet matching between all fab metrology tools — integrated or standalone — becomes mandatory. Moreover, as the transfer of process technology between sites and companies becomes commonplace, worldwide fleet matching of metrology tools is also required.

New materials are a challenge. As the industry learns how to control their properties, new parameters, such as crystal phase, orientation and grain size, are measured and correlated with process control steps. For APC implementation in these process areas, metrology technologies like X-ray diffraction (XRD) must move from R&D to the significantly more demanding production environment.

To do APC four to five years from now, a deeper understanding of new materials and how to control them is necessary. Fortunately, there appear to be no fundamental barriers for optical techniques, such as scatterometry down to 22 nm. In some instances, X-ray will replace optical measurements.

Jackie Seto, managing director of software, MEMS and 3-D IC at Lam Research Corp. (Fremont, Calif.), views APC/AEC as crucial. "We're seeing instances where process requirements exceed equipment capabilities," she said. "There will always be natural equipment variations due to things as minor as tolerance stacking of pieces composing the equipment. APC/AEC can minimize this. It'll be more cost-effective to use software and controls to ameliorate those variations' effects on the output."

Software is increasing in importance because limitations are being reached on what systems can do. This is evident in lithography. The industry must tighten equipment performance using APC/AEC to get what is needed using software control, rather than error-prone human intervention. This will be done through feedback loops — methods of automating calibrations making them system-to-system consistent — to better match up chambers. This will require additional and new sensors. The key to this kind of APC/AEC will be having the right sensors and the data quality required to get the necessary high precision.

Embedded knowledge is the trend for APC/AEC software; software will be designed to create expert, smarter systems. Having platforms perform the same — chamber-to-chamber, system-to-system and fab-to-fab — around the world will hinge on how consistently these chambers can be made to perform. One way is to simplify how things are maintained, analyzed and troubleshot by embedding more knowledge into the systems.

"APC/AEC software is limited by, and is as good as, the sensors," Seto said. "If something is undetected, the software is irrelevant. If the sensor technology is available, another challenge is what the software is saying — the correlation of information coming from the chamber vs. what takes place on the wafer. For years, there has been talk about doing multi-variant fault detection. For it to be effective, there must be a direct correlation between the series of incidents being detected in a multi-variant situation, and a specific incident on the wafer that is known will cause a yield, performance or scrap issue." A higher level of mathematical skills will be required to produce the algorithms to understand platform data and its correlation to wafer performance. Accurate data collection will be necessary. If the data to support the qualification process is unavailable, then decisions will be made based on past experience on how to modify a process to hit the target.

The OEM may have different customers who require different data sets, and must solve how to get the needed data available at the right time while isolating the different sets of data needed by different users. This is complicated by the fact that the OEM sometimes considers data from the equipment as intellectual property (IP) that must be protected. Some of these IP concepts will have to change. To meet yield requirements, the OEM will have to provide a richer data set, as a whole new level of control data will be needed.