Defect Detection Overcomes Limitations

High-aspect ratio inspection is a major challenge to defect inspection and review tools. Damascene interconnect structures demand extraordinary inspection capabilities. While horizontal features are small, perhaps 0.18 µm, some depths can be at a more than 8:1 ratio, ~1.5 µm deep! Pumping sufficient energy into those narrow, deep features to perform a proper inspection is an extreme requirement.

Many fabs that have invested in CMP and lithography tools for fast shrinks have not made the necessary supporting defect inspection investment in the litho cell (Fig. 3). As a sales engineer of one of the major equipment manufacturers put it, 'Litho modules sometimes operate as if in a vacuum. If you're doing three shrinks a year, what you're talking about is litho cell changes. To succeed, you'd better have a well thought out litho inspection and yield management strategy. This is the cosmic 'Duh!' I've had customers tell us, 'We didn't have to do this before!' and I feel like grabbing them by the ears and shouting, 'You never did three shrinks in one year before either!''

Many fabs are not increasing their yield management staff, adding to quick technology adoption difficulties. This is why some equipment providers offer soup-to-nuts yield management consulting services tailored to the fab's operation. Although not yet a fully adopted strategy, many fabs are resorting to these services to accelerate yield management learning and control costs.

Fig. 3. Litho defect after etch on an advanced logic device. An investment in the latest CMP and lithography tools must be supported with necessary defect inspection systems; otherwise doing three shrinks a year can become a risky proposition. (Source: Applied Materials)

As the same sales engineer put it, 'Lots of people seem to be running on hope. They do shrinks working from the premise that 'I won't worry unless we get a yield smack.' If it happens, it is often catastrophic to yields and finances. Smart manufacturers aren't in denial and invest now to figure out how to avoid problems before they occur. Even during the downturn, they're developing their yield management capabilities and will have a competitive advantage when the upturn comes.'

Data: How much is enough?

When dealing with complicated processes, such as copper, fabs must determine how to engineer each process step to best benefit the following step and the next and so on. It is difficult to determine where to optimize by just looking at individual process steps. When these steps come together, and their interactions are studied, benefits accrue, because equipment can be fine-tuned, inspection strategies changed and the overall processing scheme altered.

This is why Applied Materials (Santa Clara, Calif.) created its Equipment and Process Integration Center (EPIC), housing equipment and processes needed for fabs to develop and test a completely integrated multilevel metal copper interconnect process before actual equipment installation.

Pat Lamey, Applied's senior director of Wafer Defect Inspection Systems, Process Diagnostic and Control Group, views metrology as an important aspect. 'At EPIC we have wafer inspection tools, a SEM review tool, as well as a defect data management system that talks to the inspection stations. It tells us what defect density was when it went through a process point, what the results were after the inspection, and the little black dots got categorized. It looks at the various tools (not just ours, obviously) and presents the customer with reports detailing what is known about process performance and defect densities, types, mechanism and sources.'

In developing a copper layer, Lamey believes it is not always necessary to determine how to configure the inspection system to react to it. 'Now we can go to the engineers putting the process together and ask, for example, whether the copper layer need be as bright as it is and if this might be changed without affecting electrical performance. Or, if a material has too much grain, query whether grain size can be controlled to reduce noise from the underlying pattern, improving defect capture chances without affecting material performance, either for its etchability, defect susceptibility or, most importantly, electrical characteristics.'

Higher inspection sensitivity is needed, because smaller dimensions and new materials are creating images never seen before. Inspection tools must deal with features a fraction of the wavelength of light! Alternatives for visible light, such as laser scattering, will be needed. This is a migration from image processing into signal processing. Today, light can show that something has changed, although it may not be able to show what it was that changed. This is sufficient to bring SEM tools to bear to identify the problem. Thus, laser scattering ­ which provides insight into pattern perturbations ­ with SEM technology speeds up what was a slow full-wafer inspection procedure.

However, now it becomes possible to find defects that do not compromise the die. Lamey recalled a demonstration done for a user who wanted to see how sensitive the tools in his facility were. 'What was nominally a production recipe that detected 100-300 defects (of interest) per wafer, when run with maximum effort, produced 11,000 defect events. The customer agreed that the first 200 he reviewed were very real events. But were they electrical faults? Absolutely not!'

The next step is what to do with this information. 'Currently,' Lamey said, 'on-the-fly ADC is used. When the user gets 11,000 events, the system indicates that 10,792 are defects they don't care about.' Because this is signal processing, it is possible to monitor multiple information channels to make determinations about the detected events. 'I've seen results where the system shows it has found 125 defects, simultaneously categorizing 42 of them as a large scratch, 22 as microscratches and the rest as surface particles. If the user doesn't care about surface particles but cares about microscratches, the system automatically reviews those 22 defects deemed important, bypassing the rest. The user sets the bar.'

Lamey believes some fabs tend to implement defect strategies that result in gigabytes of data. 'The amount of information is less important than how to get it to tell you what's wrong and how to fix it.' Lamey said many fabs accumulate large quantities of defect data, because they like the idea that they can perform data mining and analysis, produce projections and predictions and confirm that was has been done is right or not. 'But they don't focus on what to do to make it better now,' he added.

Defect libraries must be revised. New processes are creating new defect categories. Some events taking place in copper have never been seen on aluminum. Dielectric defect mechanisms can result from an interaction of the copper with the chemistry. This is a challenge, because until the electrical test is carried out, it cannot be determined whether a visual defect is an electrical defect.

Automation capabilities, in terms of defect classification and managing data management systems, are undergoing extensive refinement and should soon surpass high-resolution defect classification, which is used by many fabs to replace manual classification techniques. 'When we move our present capability to the next plateau, we will be able to do 95% accuracy and purity on automated classification of defects as opposed to today's 85%,' Lamey said.

Defect detection and voltage contrast

Traditionally, inspection technology relied on optical techniques to perform in- and off-line defect detection in semiconductor manufacturing. Now, an increasing proportion of defects is optically undetectable. Additionally, smaller features can make some optical systems produce high nuisance defect rates.

About 50% of 0.25 µm and below feature defects are optically undetectable. This is partly due to smaller geometries, but the major barrier is subsurface defects in interconnect layers. Optical techniques lacking electrical insight into what they look at cannot detect subsurface problems.

Schlumberger ATE (San Jose) has added a twist to inspection by developing an application using voltage contrast technology for killer defect detection.

The technology detects differences in charge (or voltage, hence image contrast) between floating conductors on one area and grounded conductors in an equivalent area. By comparing contrast levels, electrical subsurface defects such as opens and shorts can be found, which are otherwise optically undetectable. An advantage of the technique is that nuisance rates are inherently low and correlation to electrical inherently high.

According to Andy Pindar, vice president of the Diagnostics Group, 'E-beam use isn't limited to voltage contrast. High detection rates with pattern defects are also important to users, as technologies shrink and optical tool sensitivity runs out of steam.' The system enables detection of open contacts, shorts from stringers, particles that might cause a short between two conductors, cracks in a conductor, under- or over-etched contacts and other potential problems that optical systems are either blind to, or have low detection capability for.

Chris Talbot, vice president of technology and strategic marketing for the Diagnostic Systems group views copper as a major application for the system. 'The dual damascene structure's high-aspect ratio makes detecting optically anything at the bottom of either the damascene trench or the underlying via contact extremely difficult.'

The system is optimized to make voltage contrast production-worthy. Toolboxes for fine-tuning voltage contrast and conventional E-beam imaging provide flexibility in throughput and sensitivity selection. The obstacle was speed (throughput, area coverage rate), a disadvantage with E-beam-based defect detection tools, which are significantly slower than optical tools.

Another concern was flexibility. As Talbot put it, 'If you're going to have a tool that doesn't do wafers per hour, then you need a flexible architecture that allows the use of intelligent sampling scenarios to compare any piece of any die to any piece of any other die to locate and characterize systematic defect signatures.'

Schlumberger is the first to admit that this is a complementary method, and that other defects require different approaches. E-beam does not make optical inspection tools obsolete. There will always be a need for a bolder detector to detect gross equipment problems and particle types of defects.

Pindar predicts a move away from use of one or two tools to detect everything throughout the process. 'There will be tools optimized for specific defects found at particular process steps,' he said. 'Voltage contrast defect detection will be one of an array of detection technologies applied to the interconnect and via layers.'

This development would appear to be emerging at an opportune time. As major players in the industry push from 0.18 to 0.13 µm, their biggest concerns are subsurface electrical defects in interconnect layers. Finding these at final testing is an unsatisfactory solution.

The system's detection range is considerable. It can detect an open unfilled via or contact, an open filled via or contact, and it opens in metal interconnect. The possibility of opens in copper dual damascene structures, caused by incomplete etch at a via hole's bottom, is a concern; however, voltage contrast techniques have proven successful in early copper applications. In the shorts category, the system can detect gate oxide shorts and shorts in salicide over the poly layer. Although a complicated process, it can also find metal layer shorts. Schlumberger believes its system can reliably detect over 80% of the killer, optically undetectable, voltage contrast defects with a very low nuisance rate.

The technique is not a cure-all. Although it works well, in some tight arrays, such as memories, there may be too much leakage current for it to work, requiring standard SEM imaging to provide the solution, where optical is affected by scattering. Also, although when using voltage contrast to detect a clear signature, nuisance defect rates plummet, when looking at individual pixels, it rises as it would for an optical system. Some killer defects do not exhibit voltage contrast signatures. Their detection, difficult with optical means, is important to identify yield failures and implement process fixes. For example, a source/drain, bridge or break will not show voltage contrast effects, due to direct contact with silicon, requiring other methods.

The system is expected to be available during the second half of the year.

Determining defect composition

Fred Stevie, distinguished member, technical staff, at Lucent Technologies' Cirent Semiconductor Analytical Laboratory (Orlando, Fla.) still views determining elemental composition as a major defect inspection and review challenge. 'Unless you know what that defect or particle is made of, you're uncertain how to go after it; there are too many steps, too many things to look at. Once identified, you know what the probable solution is.'

Auger electron spectroscopy is a proven method for identifying surface defects, but it cannot answer all questions if the defect is below the surface, especially if that surface is an insulator.

'A commonly used method is energy-dispersive X-ray on SEMs,' Stevie said. 'The problem is that, with a very small defect or particle, the information zone from where your X-rays emanate is too deep, and although you're getting information on the particle, you're getting far more from behind the particle. That has limited SEM capabilities using EDS.'

Stevie sees TEM as the technique of choice. 'With a high-resolution TEM, the specimen is very thin. When you use energy-dispersive X-ray methods, the region that complicated the analysis before is gone.' Stevie added that since it is possible to obtain elemental information from an area as small as 10 Å, given the SIA Roadmap's requirement for detection of 300 Å particles over the next few years, the tool to do it with exists now.

Stevie believes that secondary ion mass spectrometry, which offers high sensitivity and considerable depth resolution, will become a standard inspection tool when used with FIB sources, which have excellent lateral resolution. The limiting factor is the need to improve secondary ion yields.

Optics ­ bloodied but unbowed

Attempts to hammer a wooden stake into optical inspection's heart are so far unsuccessful. According to Werner Hunn, director of product and market development at Leica (Allendale, N.J.), 'With the coming of higher requirements for defect recognition and quality control, optical inspection stations are also migrating into, for example, back-end applications such as bonding pad inspection and defect analysis in the back-end stages after electrical testing or even after the passivation layers and dicing.'

As geometries (and defects) become smaller, optical tool providers such as Leica, Zeiss (Thornwood, N.Y.) and Nikon (Melville, N.Y.) have extended their systems' sensitivity into UV and DUV ranges, adding laser-controlled scanning (Fig. 4). Optical inspection offers a COO advantage in a field related to defect review, where defects detected by high-speed systems have to be reviewed, automatically analyzed and classified.

'Optical technology still has a long way to go,' Hunn said. 'With I-line UV, we can detect particulates down to 0.13 µm, 0.10 µm with enhancement. DUV goes even beyond. Certainly, techniques such as E-beam technology will be tried out as well as AFM, but optical technology will still be in use as it migrates to lower wavelengths and higher automation. It will be active in automated defect analysis and image archiving of critical defects.'

Optics' limitations are exaggerated. As Hunn put it, 'The perception is that nothing useful can be seen below 0.5 µm. That's not true. We overemphasize small defects in processing.'

From a dimensional point, only a few layers are critical. Other conditions exist on a macro scale. In the photoresist process, for example, many problems result in macrodefects. 'We've seen a strong interest in coupling on the wafer stations the visual macro inspect with the micro inspect,' Hunn said. 'And micro inspection is not necessarily done at the highest possible optical magnifications. There are inspection and sampling techniques where only a 20X or 50X objective is used.' In other areas, lithography for example, high-magnification optical inspection will continue playing a crucial role.

Claus Nielsen, marketing manager of August Technology (Edina, Minn.), agrees and looks forward to users developing a better understanding of this inspection area. 'There appears to be a disparity between what defect inspection tool users think they need to find and the actual defects they want to find once they develop a better understanding of the automated back-end wafer inspection process. The traditional manual inspection process has led some to think they're finding defects that are smaller and more critical than they truly are.' According to Nielsen, few companies have solid data on the percentage of failure reductions based on discarding die based on certain visual qualities.

Moore's Law's long reach

Smaller defect detection requirements not only affect detection instruments, but also their mechanical components, including stage navigation accuracy. As James Jackman, vice president of Strategic Marketing for FEI Co. (Hillsboro, Ore.), put it, 'Over the last two years the killer defect sizes have been halved, requiring twice the magnification to see them. Available fields of view become smaller, requiring twice the overall positional accuracy to ensure that the defect can be found.'

FEI, like others, is raising stage accuracies to reduce defect location uncertainty. 'We still make a contribution,' Jackman conceded, 'but as we have a less than 2 µm error, and defect detection tools typically have a 4 µm error, using RMS addition you arrive at a total inaccuracy of about 4.5 µm. That means we've only added 0.5 µm to it.'

Jackman views the move from 200 to 300 mm with concern. 'We'll need things like wafer chuck temperature stabilization, because otherwise thermal-induced expansion could significantly impact the accuracy of the defect 'find.' Automation is the answer; being able to automatically realign to the die origin will improve accuracy.'

Things to come

Ongoing trends will continue. There will be a significantly faster than predicted deployment of SEM inspection. The modularization of inspection, classification and analysis will continue, tying them closer together. All this will be driven by shorter device generation lifespans, leading to new defect types and the need to understand quickly these defect populations. Comprehensive integration across different inspection and metrology tools will be increasingly critical.

Currently, the industry operates on the diminishing returns side of many of the technologies in place, which are stretched to their limits. Although we are accustomed to pushing the envelope, issues that we have never had to cope with loom on the horizon. For one, we are approaching actual physical limits. The SIA Roadmap itself is in tatters; two years back, nobody would have predicted we would be talking about SEM inspection now. With most toolmakers' demo capacity at 0.13 µm, nobody doubts defect detection and review will get tougher.