VIEWPOINT

Assessing future trends in automated defect inspection and yield management

(Second of two parts)

Paul Sandland

The semiconductor industry's push toward smaller geometries and the use of novel processes and new materials inevitably leads to a less stable manufacturing process which requires tighter closed-loop automated control than ever before. Not only must major yield excursions be detected and corrected, but yield levels must be raised through continuous yet discrete improvements to the process and equipment.

To accomplish this, critical data have to be intelligently, quickly, and efficiently acquired and analyzed. But increasing the sampling rate and defect sensitivity of inspection systems—as well as determining more intelligent methods of sampling—is not enough. Inspection, defect classification, and yield analysis systems must be more automated and integrated in order to shorten the time needed to identify causes of yield loss and, preferably, target a particular tool that requires maintenance or a process that needs improvement.

Current inspection equipment can only detect major yield excursions that need a small sample of defect data and track slow background level improvement when a long time can be taken to collect a statistically confident sample. The challenge for future inspection and yield management solutions will be to produce predictive systems that have enough sensitivity and repeatability to prevent defect excursions by automatically alerting the user to potential problems or modifying equipment parameters before significant yield loss occurs. To effectively accomplish this, inspection, automatic defect classification, and yield analysis systems must be closely integrated into an overall yield management solution. Linking these elements through intelligent software algorithms will result in a system with close interactions that will provide users with immediate, powerful data feedback.

Future Inspection Systems

The process of defect inspection can be seen as the passage of a very large amount of data through a series of sieves to get a one-line answer: "tool or process X is producing excess defects." While the speed of defect detection must continue to be increased, the ultimate goal is the minimization of the total time to detection and correction of a problem.

Continuous yield improvement (or yield learning) and yield excursion detection are critical elements of the yield management process, but each requires different inspection techniques to be effective. Yield learning calls for an inspection system that can detect all critical defect types within a process module. Defect type and size sensitivity are the most crucial requirements since a high defect capture rate is the final objective. Overall throughput, though still important, is a secondary element because all defect contributions to the module in question must be well understood to achieve baseline yield improvement. For defect excursion monitoring, the capture rate of critical defect types is again important, but throughput becomes more critical since the goal is a quick determination of whether a process or tool is out of control. The manufacturer can then protect those products that would be at risk if the defect excursion were not detected.

In terms of inspection tools, yield learning applications are generally better served by optical imaging or scanning electron microscope—based inspection tools, while laser scanning tools have distinct advantages for defect excursion or station monitoring. However, since these systems are based on different technologies, they do have some distinct advantages for detecting certain defect types, and equipment suppliers can help users analyze their tool selections for these applications.

In addition to having the right combination of sensitivity and throughput, inspection tools will also need to be closely coupled with automatic defect classification and yield analysis systems. Linking these tools into an integrated package will provide users with additional power for solving yield issues quickly and efficiently.

Automatic Defect Classification

An emerging technology that is already beginning to have a huge impact on yield enhancement is automatic defect classification (ADC). ADC replaces manual microscope classification inspections performed by engineers and technicians, providing higher performance in classification accuracy and repeatability (by reducing subjectivity and classification errors) and significant increases in overall throughput. ADC can be done either during first-pass inspection, through a second-pass classification on an inspection system, or off-line on a review station.

The more information one can gather about a defect, the more likely that accurate and useful classification can be attained, leading to the isolation of the defect cause. The spatial distribution signature is often all that is needed to identify the root cause. In other cases, a high-resolution optical image renders enough information to detect the defect source. Sometimes it is not until the defect has been imaged using an electron microscope or analyzed using Auger spectrometry to detect chemical composition that the defect cause can be ascertained.

The defect finder can be run at the smallest pixel size (i.e., highest resolution) and gather data for defect classification on the fly, or to run at a lower resolution in a shorter time and use a second pass at maximum resolution to classify the defects. Generally, defect detection systems are operated at less than maximum resolution because of the need to cover enough area to gather a statistically valid defect count in a short time. Since for any defect class it is only necessary to visit a statistically significant sample, it is often faster to perform two passes than one. Another advantage of a second pass is that it helps to eliminate noise since each potential site is visited twice. In order to be consistent with meeting the speed requirements, all possible information should be extracted from the first-pass image.

Another ADC application is the global or macro identification of wafer pattern signatures (see Figure 1). A spatial analysis of defect wafer maps can provide clues to the defect sources. Defect excursions can produce defect clusters that have a distinct spatial signature, and these signatures can be tied to particular equipment faults. By using an algorithm that recognizes these characteristic patterns and tracking the path of product through particular process equipment, spatial signature analysis can identify tools that need to be shut down for adjustments or cleaned.

Yield Analysis

Two main tasks of a successful defect data management and yield analysis system are yield/defect correlations and statistical monitoring of the process line. Other critical benefits that such a system can supply are improvements to both inspector performance and engineering productivity levels.

Overall yield enhancement productivity can be improved through optimal sampling plans. A simple, fixed sampling plan of, say, two wafers per lot may not provide results that have appropriate statistical confidence, particularly as yields improve. This may lead to missed yield loss excursions or an increase in the time it takes to raise the baseline yield levels of a particular process. At the same time, too large a sampling plan means that system utilization is not optimal. Defect issues that require more inspection time to detect may be missed because of poor system utilization.

The more information one can gather about a defect, the more likely that accurate and useful classification can be obtained, leading to the isolation of the defect cause. Photo Courtesy of KLA-TENCOR.

By using a database of information and intelligent algorithms, an expert yield analysis system can help provide closed-loop adaptive inspection sampling that will maximize inspection system efficiency as well as the overall yield management system. This effectively improves system utilization, lowers cost of ownership, increases overall equipment effectiveness, and enables the user to detect defect/yield events faster, thus raising overall productivity.

The identification and elimination of yield-limiting defects is a time-consuming process that requires significant focus and resources. Automating this process would improve both the time it takes to achieve results and overall productivity. Through automatic identification of yield-limiting defects, end-users will be able to concentrate their resources quickly and effectively on critical yield-affecting issues. The automation of production reports and integration of numerous data types into common databases for analysis would also save engineers and technicians hours of tedious work and dramatically improve their productivity.

Automatic identification of yield-limiting defects requires the automation of yield correlation and statistical tracking. Systems must be able to automate yield correlation through the integration of inspection, classification, process flow, and electrical test databases, coupled with automated analysis of electrical map and defect map overlays, defect signature analysis, and other advanced analysis techniques. Automated analysis of this information will allow yield enhancement teams to receive yield statistics by process level and defect types. These teams will then have the ability to zero in quickly on defect types and process layers that most affect yield. Eventually this will lead to direct identification of—and interactions with—process modules or equipment causing yield loss.

Figure 1: One ADC application is the global or macro identification of wafer pattern signatures.

Yield analysis systems also need to continue to track defect data with statistical process control monitoring but they must do it more intelligently. By having access to a tremendous amount of data, analysis systems will have to provide alternative methods to tracking defects so that excursions are not hidden within the baseline defectivity noise of a process module. For example, ADC provides the analysis system with the capability of statistically tracking only fatal defects by classification type. A specific fatal defect type may reach excursion levels—and high yield loss—at relatively low defect counts. If a wafer has a large number of nonfatal defects or a long scratch (which produces thousands of defect counts), this excursion can be lost in the noise if the total defect count was the only metric being tracked statistically. An intelligent yield analysis system should automatically search for these trends and highlight options for statistical tracking to the user.

Conclusion

An ideal inspection system would detect the precursors to yield loss and provide corrections in a short time compared to the rate of yield change. To do this economically, we need to identify the temporal defect signatures of fab equipment and their relationship to yield loss so that cost-effective defect inspection systems can be built that only seek specific problems at specific times. Slowing down the rate of change of performance of production equipment will allow longer inspection times. New solutions will be required to provide cost-effective defect detection, classification, and control for 100-nm and finer geometries. Defect detection will continue to use a mix of technologies, such as bright-field imaging at shorter wavelengths, dark-field and Fourier filter techniques, and more high-speed electron microscopy. No single technology will become dominant because each method has strengths and weaknesses. Instead, a strategy of multiple inspection technologies combined with an advanced line monitoring and yield analysis system will likely be the industry's chosen approach to the myriad of future inspection and yield challenges.

Acknowledgments

This article was adapted from a presentation titled "Automated Defect Inspection: Past, Present, and Future," originally presented at SPIE Microlithography, Santa Clara, CA, February 1998. The author would like to thank Robert Cappel of KLA-Tencor for his help in preparing this article. (The first part of this article appeared in the June 1998 issue of MICRO.)

Paul Sandland is a senior technical adviser for KLA-Tencor Corp., San Jose. Since joining KLA in 1976, he has been instrumental in the conception, design, and development of numerous inspection technologies. He holds five patents, has written many articles and papers, and has been the recipient of several industry honors and awards.

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