ANALYSIS AND METROLOGY—Defect Classification

Using 'on-the-fly' automatic defect classification to enhance yields

Andy Skumanich, Applied Materials PDC Group; and Reinhold Ott, White Oak Semiconductor

Since development cycle times have continued to shrink as outlined in the National Technology Roadmap for Semiconductors, it has become increasingly important to identify and isolate wafer defects in order to meet yield targets and maintain profitability. Real-time defect capture alone will not improve yield, however; it is necessary to characterize defects and trace them to their sources so that problems with processing tools can be corrected. The more defect information that is quickly available, the more likely it is that appropriate action will be taken expeditiously, enhancing yield and productivity. Not only can rapid defect-source identification prevent the continued use of lots in a problem process, but it can also help fabs avoid shutting down lines unnecessarily. Moreover, when a problem is flagged, defect-source identification can be useful in bringing a system back up as quickly as possible. Thus, effective yield management requires both reliable defect information and rapid defect-source identification.

Traditionally, defect reduction and yield enhancement strategies have followed a set sequence of events. Defects are captured and located with a wafer inspection system. Then a small subset of defects is subjected to manual or off-line review for defect-type determination. In the next step, a defect Pareto analysis indicating the respective counts for the various defect types is generated. The defects are then analyzed in more detail to determine potential defect sources. Finally, defect source parameters are evaluated and corrective action is taken.

Besides being time-consuming, this approach does not guarantee that defect samples will represent the entire population. This article describes the implementation at White Oak Semiconductor (Richmond, VA) of a wafer inspection system that can classify and count defects in-line with no impact on throughput. Known as on-the-fly automatic defect classification (OTF-ADC), this technique, developed by Applied Materials (Santa Clara, CA), provides immediate defect segregation. Also, since some defect categories can be tied to their most likely causes, this system assists in defect sourcing.

OTF-ADC benefits both excursion monitoring and process development. When it is used in-line in conjunction with prior defect analyses, defect-generating problems can be identified more rapidly than with a standard total defect count and review procedure. Certain defect types that may not be relevant at a given process step can still contribute to the total defect count. If the operator monitors only total counts, variations in critical defect types can remain hidden. OTF-ADC segregates defect categories so that yield-limiting defects can be monitored independently. In this manner, key yield information is readily available and the time needed for source identification can be reduced.

ADC Concept

As outlined above, wafer inspection tools historically have performed only the first step of a defect reduction and yield improvement process—defect capture. Following capture, the types of defects must be identified and an operator must analyze a subset of defects to evaluate their possible origins. Although some automated classification tools are available, the identification and analysis steps are frequently time-consuming and require a high level of user expertise.

To perform the identification step, operators must revisit defects at manually selected sites using either an inspection tool, a stand-alone optical review station, or a scanning electron microscope (SEM). The additional image information gained from these procedures enables them to perform defect typing. However, because these tools are capable of investigating only selected sites, critical defects can be overlooked.

Figure 1: Schematic showing the multiperspective architecture of the OTF-ADC system.

The on-the-fly ADC technology was developed to overcome these problems by using wafer inspection to acquire defect characterization information. Because of the multiperspective architecture of the OTF-ADC system, different defects have different optical scattering signatures, which are registered by multiple detectors positioned around the wafer, as shown in Figure 1.^1,2 The probing light beam strikes the wafer surface from above. Then the beam and wafer stage are translated to cover all the dies on the wafer. Dark- and bright-field measurements are taken simultaneously. The resulting defect data are automatically segregated into a few relevant categories, after which the defect density for each category is quantified and the output is used for further analysis. In addition, the technology's architecture permits good defect capture, because it is effectively immune to CMP "color noise."³

The ADC system was designed to provide useful data, to be flexible, and to be easy to use. The data collected through this method are useful in that they provide information relevant to specific processes, particularly those that generate killer defects. The system's flexibility enables technicians to inspect and categorize various layers and defect types. Finally, the system requires only a minimum of "teaching," so that after tuning the recipe it is not necessary to provide numerous defect examples to define the classifications.

	Inspection Method
Capability	Stand-Alone	Revisit	On-the-Fly
Usage	Off-line	On-line	Concurrent
Defects processed	Subset	Subset	100%
Impact on throughput	High	Moderate	None
Classification sensitivity	Moderate	Moderate	Low
Extra handling	Significant	Minimal	None
Redetection	Yes	Yes	No
Footprint	Increased	Same	Same
Turnaround time	Significant	High	None
Tuning learning	Classifiers	Classifiers	Thresholds

Table I: Comparison between the OTF-ADC system and standard methods of defect classification.

Table I compares the various approaches to defect classification. While OTF-ADC is the only technique that does not affect wafer throughput, its extent of defect categorization sensitivity is reduced. Typically, only a few defect categories are selected for analysis.

System Evaluation

The OTF-ADC system has been used at White Oak Semiconductor for both excursion monitoring and process development (for example, baseline definition and process learning). In the former case, the defects of interest are monitored, and in the latter, further analyses are based on the "smart sampling" of the defects. Because the OTF-ADC system categorizes 100% of the defects, the yield manager uses the defect segregation to determine whether further review is necessary and, if so, which defects should be addressed.⁴ Detailed analyses are then performed using an automated SEM, which accepts output from the OTF-ADC tool.

In White Oak's initial study, the OTF-ADC system's capabilities were evaluated at the post—trench etch, post—gate etch, and post-CMP layers. A set of 10 wafers was inspected at each layer. The inspection tool was configured to capture and classify defects into two categories at each level. Only initial tunings were used, which were not fully optimized. The key parameters for evaluating the system's effectiveness were accuracy (the ratio of the number of classified defects detected to the actual number of defects of that type), purity (the ratio of correct classifications to the total number of defects detected in that class), and repeatability (the wafer-to-wafer comparison, whose formula is R = 1 — 3 […/mean]). For example, in a hypothetical case of a wafer with 50 scratches, the OTF-ADC system may classify 40 defects as scratches. If an operator were to review the defects and find that 10 of those 40 were particles, OTF/ADC would have an accuracy of 60% (30 of 50) and a purity of 75% (30 of 40).

For both etch layers studied, pattern defects are the most important because they reflect process issues; generally, particle defects can be removed and are less critical. For the CMP layer, microscratches are the defects of interest, reflecting the status of specific process conditions. Therefore, for the postetch layers, OTF-ADC was set to separate defects into the two categories of particles and pattern defects, and for the post-CMP inspection, the categories used were scratches and "other defects." Tables II and III summarize the results for these defect categories and the three key ADC parameters (accuracy, purity, and repeatability). In the cases studied, all of the classifications showed high accuracy and very good repeatability (78%). The primary defect classifications showed excellent purity, especially those at the post-CMP inspection phase.

Etch Process Step	Pattern		Particle		Repeatability
	Accuracy	Purity	Accuracy	Purity
Post—trench etch	75%	92%	71%	38%	—
Post—gate etch	75%	97%	90%	47%	78%

Table II: OTF-ADC results for postetch inspections.

Nonscratch Defects		Microscratches
Accuracy	Purity	Accuracy	Purity
97%	98%	89%	93%

Table III: OTF-ADC results for the post-CMP inspection.

The effectiveness of the line monitoring of defect types depends on high accuracy and purity of the defect classification system. In this initial study the system met these criteria for a key defect type for each layer. For the two etch layers, pattern defects were captured with high accuracy (75%) and purity (>92%), while for the post-CMP layer accuracy and purity for scratches were 89% and 93%, respectively. These high levels of accuracy and purity, as well as repeatability, indicate that the defect segregation capabilities of the OTF-ADC system can be used with confidence to inspect these layers. If further studies are necessary, a SEM can be utilized for more detailed characterization of specific defects. In Figure 2 some representative SEM micrographs of defects captured and categorized by OTF-ADC are shown. In addition, the system is being refined to increase purity results for postetch particles. While defects caused by postetch particles are not of major interest for the excursion monitoring of the two etch layers and the post-CMP layer, increased purity will facilitate process development applications where all defect types are studied. Even at current purity levels, the high accuracy of OTF-ADC assures that particular defects can be categorized properly and that when technicians strive to select several samples of a given defect for review, at least some of them will be included.

Figure 2: SEM micrographs of defects categorized by the OTF-ADC system: (a) pattern defect, (b) particle defect, and (c) microscratch defect.

Implementation Benefits

Before shutting down a tool or module to correct a defect-generating problem, an accurate defect classification system must be in place. Given the OTF-ADC system's accuracy and purity results for CMP microscratch classifications, the system was set up at White Oak for post-CMP monitoring. It replaced a classification procedure in which between 10 and 20 defects (~10% of the total detected during wafer inspection) were reviewed manually. An example of the data provided by the system is presented in Figure 3, which shows microscratch and total defect counts as a function of lot. Looking at the total defect count does not give a true indication of the microscratch-count excursion. Only segregation conclusively indicates when microscratches are out of control. Because this defect excursion was captured in the microscratch monitoring data, the problem source could be determined rapidly. The microscratch data were useful because they tied the defects to their source.

Figure 3: Typical results of post-CMP inspections in which counts are shown as a function of lot: (a) total defect counts and (b) microscratch defect counts.

The separate monitoring of microscratch defects has led to a significant decrease in the time required to detect excursions for the post-CMP process step. Figure 4 compares the time required to perform excursion detection with classified defect counts and with total counts. The variation window is smaller for the classified counts. Consequently, if the microscratch counts exceed the alarm threshold, this variation will be detected sooner than will a variation in total-count data, in which several defects contribute to the variation window. A time savings of as much as a half day, or six to eight lots under the given process conditions, has been observed. In a full production mode, the number of lots affected could be substantially higher, representing an even greater potential cost savings.

Figure 4: Comparison between the time required to detect defect excursions using classified (microscratch) defect counts and total defect counts.

There are advantages to using the defect classification system during process development. The classified counts can be useful for both baselining and excursion evaluation because the decision path for defect analysis depends on which type of defect is increasing. Figure 5 outlines how using the OTF-ADC system can reduce uncertainty in defect analysis and source identification. The most certain method of detecting the source of defects is OTF-ADC, because this system reveals more about defects and their potential causes than conventional detection systems.

Figure 5: Comparison between the certainty achievable in defect source identification using conventional and OTF-ADC inspection systems (YM=yield management).

When conventional total-count systems are used and more than 300 defects are detected, the operator must analyze a sample defect subset to determine the possible problem source. Although theoretically random, the sample subset may be influenced by operator bias and may not represent the entire defect population. Moreover, if an insufficient number of defects is chosen, the review process might need to be repeated. In contrast, when OTF-ADC is used for microscratch detection, there are two excursion limits. If the microscratch count is low while the count of other defects is above the alarm level of 200, the technician can confidently select a representative set of nonscratch defects for further analysis. The possible sources have already been narrowed because the scratches are not involved. If only the microscratch count is above the alarm level of 100, the likely defect sources are already known and further defect analysis is not necessary. In other words, classified-count data provide more direct information than total-count data, which reduces the need for or obviates further analysis.

Conclusion

Because numerous defect sources can cause yield loss in advanced semiconductor processing, it is essential to obtain detailed defect information in the shortest possible time with the highest level of accuracy. Historically, wafer inspection tools have provided minimal information about defects, such as counts, location, and—in some cases—size. With inspection systems capable of classifying 100% of the defects detected into a select group of categories that can be linked to specific defect sources, corrective action can be taken more rapidly than previously possible, thereby enhancing yield. Given the correlation between high yield and profitability, wafer inspection tools providing classified defect counts are likely to become an integral part of process development and monitoring activities in semiconductor fabs.

Acknowledgment

The authors appreciate the valuable wafer inspection assistance provided by Niv Ben-Mordechai of Applied Materials and the support provided by the defect reduction group of White Oak Semiconductor.

References

1. M Altamirano and A Skumanich, "Enhanced Defect Detection Capability Using Combined Brightfield/Darkfield Imaging," in Proceedings of the SPIE Conference on Microelectronic Manufacturing (Bellingham, WA: International Society for Optical Engineering, 1998), 60—64.

2. BM Nebeker and D Hirleman, "Prediction of Light Scattering Characteristics of Particles and Structures on Surfaces by the Coupled-Dipole Method," in Proceedings of the SPIE Conference on Metrology, Inspection, and Process Control for Microlithography (Bellingham, WA: International Society for Optical Engineering, 1996), 690—697.

3. A Skumanich, "Advanced Wafer Defect Detection for CMP," European Semiconductor 20, no. 3 (1998): 33—36.

4. A Skumanich and M Cai, "CMP Process Development Based on Rapid Automatic Defect Classification," in Proceedings of the SPIE Conference on Microelectronic Manufacturing Technologies (Bellingham, WA: International Society for Optical Engineering, 1999), 76—88.

Andy Skumanich, PhD, is a senior technologist in the methodical defect reduction group at Applied Materials (Santa Clara, CA). He has been active in the fields of both optical and electron beam defect capture and ADC and process development. Before joining Applied, he was a consultant for thin-film and semiconductor analysis and was awarded a contract to develop a photothermal metrology system with Lockheed-Martin. He also worked at KLA-Tencor on advanced technology development for optical and electron-beam wafer inspection. Skumanich has authored papers on wafer inspection applications for defect reduction, yield enhancement, and process development and has published articles on processing, novel materials, and probe techniques. He holds several patents. A Hertz Foundation Fellow, he received his PhD in physics from the University of California at Berkeley for research on the optical and electronic properties of semiconducting thin films. (Skumanich can be reached at 408/563-2460 or Andy_Skumanich@amat.com.)

Reinhold Ott joined the ramp-up of White Oak Semiconductor (an Infineon-Motorola joint venture) in Richmond, VA, in 1997. He is responsible for wafer-level yield analysis for new products. He has implemented a state-of-the-art automatic real-time defect-yield correlation algorithm using compressed electrical bitmaps, about which the Institute of Electrical and Electronics Engineers is publishing a paper in 1999. After joining Siemens Microelectronics in Munich, Germany, in 1995, Ott worked as a development engineer in ASIC chip production. He holds two patents for automatic defect-yield correlation and defect-related yield loss prediction in DRAM fabrication. He received his diplom-ingenieur degree in physics from the Technical University of Aachen, focusing on the development of a THz emitter based on quantum Bloch oscillations in GaAs/AlGaAs superlattices. (Ott can be reached at 804/952-7941 or ottr@whiteoaksemi.com.)

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