Automatic defect classification for effective yield management
12/01/1996
Automatic defect classification for effective yield management
Louis Breaux, David Kolar, Motorola APRDL, Austin, Texas
Automatic defect classification (ADC) enables efficient process monitoring and enhances the diagnosis of process problems. ADC, which reduces large volumes of defect data to concise statements of process status, may be implemented both on-line (tightly coupled to a defect detector) and off-line (connected to a defect review tool). Although, strictly speaking, ADC is simply the automation of manual defect classification, it must be fully integrated with defect detection and analysis systems in order to realize the full benefit. An on-line implementation of ADC has been in operation in a research fab at Motorola Advanced Products Research and Development Lab (APRDL) for more than a year, and has greatly surpassed the speed and consistent accuracy capabilities of manual methods.
Smaller semiconductor production design rules, which need higher resolution inspection, have vastly increased the yield-relevant data produced by wafer inspection systems. A defect detection system connected to a defect analysis system is not a complete process control system: In order to diagnose process problems, one must group defect occurrences into types or categories. This classification then facilitates the allocation of limited resources to the most significant defect issues and allows more accurate prediction of potential yield from lots still in process. The volume of defect data now routinely produced by defect detection systems can overwhelm traditional manual classification techniques. When manual classification is the sole option, only a small sample of the defects/wafer can be classified in order to meet production time constraints.
Defects can be grouped by several different methods. For the purposes of ADC, the defects are classified based on their similarity in appearance to known, predefined categories. One therefore must view the defects in order to ascertain the pertinent visual characteristics. A different type of grouping scheme could be based entirely on the defect sizes. With the present inspection technology, such a scheme would not require any postinspection review. In fact, the ADC system under discussion can and does operate at both levels. However, most effective yield management programs involve some level of visual identification of defects.
 |
Figure 1. Active layer trend chart showing all defect types.
Typically, manual defect review is performed on off-line optical review tools or on scanning electron microscope (SEM) tools. This task is time-consuming, tedious, inconsistent, inaccurate, and costly (relative to automated classification systems). On-line ADC allows much faster classification and eliminates handling and queue time between the inspection and review tools, theoretically allowing up to 100% sampling with relatively little production impact. In practice, though, an ADC system will also be limited to a smaller sample of defects by production cycle time constraints. This paper describes an on-line automatic classification system that can:
eliminate the need for manual classification;
provide a throughput increase of approximately 3? that of manual classification;
provide a significant improvement in consistency and comparable or better accuracy relative to manual classification;
reduce handling and, therefore, contamination; and
enable the creation of defect baseline models.
At Motorola APRDL, we have installed an integrated ADC system from KLA on a KLA 2132 inspection tool at four points in our 0.25-?m development line: Active, Gate, Gate Salicide, and Metal 2. More than 1500 wafers have been processed since the installation. On average, 100 unclustered defects are classified/wafer. ADC results are studied daily by the defect reduction team and are used to prioritize defect reduction efforts.
Two types of ADC result charts are used. The first, a stacked bar chart of a given layer showing lots split up into their constituent defect categories (Fig. 1), supplies an overall picture of the fab defectivity. Figure 1 shows the Active layer trend and includes all defect types. The 11 categories represent a combination of two cluster defect types, an in-line grouping, six learned or trained defect types, an unknown or other type, and an ADC-specific type. Though the system is trained on only six of these categories, the remaining groups are a natural consequence of the hierarchical handling of the defect data.
Figure 2 is a stacked bar chart showing just the "killer" defect types identified based on yield correlation and other factors. The number of killer defect categories is greatly reduced relative to the overall Active layer trend chart. More importantly, by setting up the data analysis to key on these categories, automated process control and lot disposition can be based solely on the most significant yield-limiting defect types. The defectivity for a lot or even a layer trend can be reported automatically without the inclusion of benign defects or inspection noise (false defects).
 |
Figure 2. Active layer trend chart showing killer defect types only.
Integrated on-line ADC
The KLA ADC system has two major components, an ADC Processor and an ADC Manager. The ADC Processor, a self-contained module installed in the inspection tool, performs all ADC calculations, including clustering, in-line grouping, defect review sample selection, and high-resolution defect classification.
The ADC Manager, a stand-alone workstation, performs all setup and training functions. It communicates with multiple ADC Processors via Ethernet; only one ADC Manager is required/six inspectors.
ADC processing consists of five basic steps (Fig. 3):
1. spatial clustering,
2. in-line grouping,
3. defect review sample selection,
4. high-resolution defect image acquisition and classification, and
5. automatic transfer of classification results to an analysis system, which generates linear monitor data charts.
The first three steps occur within seconds of the completion of wafer inspection, so the delay is minimal. Note that the complete sequence requires the full integration of defect detection, review, and analysis operations.
 |
Figure 3. ADC process flow.
Clustering
Spatial clustering analysis partitions the defects into spatially clustered and randomly distributed (unclustered) components. Scratches, for instance, are fairly common defect clusters. In a typical case, one might have three clusters containing a total of 500 defects, along with a random population of 50 defects distributed across 30 die. Yet the clusters might only affect three die - the unclustered population of defects typically has a more significant impact on yield. If the percent of die affected by clusters is much less than the killer random defects, yield enhancement engineers will focus their efforts on the randomly distributed defects.
Also, since defects within any given cluster are likely to be caused by a common defect-producing mechanism, one may improve review efficiency by selecting relatively few defects from clusters. Clustering alone will typically reduce the number of candidates for defect review by 70%.
In-line grouping
After the defects have been clustered, the entire population is separated into in-line groups or bins. These groups are based on features that are extracted in real time during the inspection step (such as size, location, and brightness), as well as features that may be calculated by the inspection tool`s analysis system (such as previous vs. current layer comparisons). While these features are extracted at the inspection pixel size, and are thus heavily influenced by the inspection test itself, no additional inspection time is required.
In-line groups are defined during ADC setup. A simple but effective in-line grouping scenario might include:
one group containing random defects that are extremely small and from the current layer (i.e., candidates for SEM review);
a second group containing remaining defects from the previous inspection layers; and
a third group containing the remaining random defects from the current inspection layer.
In this case, the user would most likely be interested in selecting an automatic review sample from the third group only. These defects pertain to the current layer and are large enough for the ADC system`s optical imaging to identify.
Sample selection
Although a variety of techniques may be used to choose defects for review, the most common is simple random selection. Given a set of defined in-line groups, other methods become possible:
1. proportional sampling that selects a set of defects from each bin commensurate with that bin`s representation in the total population; or
2. prioritized sampling that selects from some bins more heavily than others, depending on the user`s assessment of the relative importance of the defects in each bin (e.g., killer vs. nonkiller types).
High-resolution defect classification
The three preceding steps - clustering, grouping, and sample selection - are almost real time and add no significant overhead to the inspection. The next ADC step is automated second pass high-resolution review. In this step, the defects selected for review are automatically "revisited": a high-resolution image is acquired for each defect in the sample. (A "reference" image from the adjacent good die may also be acquired.) This process takes approx. 2 sec/image pair. The ADC Processor software then redetects the defect in the image, extracts needed information, and performs the actual classification. The system calculates the automatic class category by comparing the extracted information to a database extracted from defects of known classification. This database contains a limited set of images selected by a human expert to represent each defect category to be trained.
The user can save images acquired during ADC operation to disk, an especially useful capability when an "out-of-control" condition has been detected. One may then review defect images from wafers immediately preceding the out-of-control condition, and compare them with defects detected on the out-of-control wafers. Or, with 100% sampling, one could obtain a set of images for the same defect showing the evolution of the defect through the process.
Analysis
Finally, the system transfers the ADC results via Ethernet to a defect analysis system. Note that at each step in the processing, the number of defects that must be processed is reduced from that of preceding steps. Thus, system performance is relatively insensitive to large changes in the number of defects encountered.
Case study - reducing the time to react to a fatal defect excursion. Figure 2 illustrates the high value that may be realized through the use of an ADC system. The system detected a killer-type defect excursion that had occurred on several lots prior to the first one reaching the inspection step several days later. After the lead lots reached the inspection point, a bar chart showing the excursion lots and their defect categories with images was available for review by the defect engineers. The defect source was quickly identified and corrective actions were taken, but several lots were scrapped. To minimize the damage in the event of a possible reoccurrence of the defect, the inspection point was moved back a number of steps, closer to the source of the problem.
Subsequently, the ADC system detected a reoccurrence of the defect. This time only one lot was affected: the defect inspection was performed on the night shift and the problem was identified and fixed the following morning before more lots were processed through the defective tool.
ADC performance
On-line ADC reduces the time needed to collect results, and the number of manufacturing steps, but these reductions would not be sufficient if the results were not as accurate as existing manual methods. Therefore, the accuracy of the ADC classification was compared with that of a fab defect expert. Calculating exact accuracy is difficult because human defect classification is subjective - it is rare for reviewers to agree on all classifications on a wafer. All automatic classifications were compared to only one human expert in order to eliminate this error source.
 |
Figure 4. ADC performance, comparing front-end and back-end accuracy for the same defect types. Overall accuracy: active - 73%; poly - 76%; salicide - 86%; metal 2 - 71%.
Other accuracy considerations should also be noted. For instance, all wafers used in this work were from normal production at 0.35 to 0.25-?m design rules. This leading-edge processing technology involves some less mature processes without the control over film thickness and uniformity that is expected for =0.5-?m design rules. These process variations are a very significant source of noise and complicate processing of the images to extract the defect features. In comparing and assessing the accuracy of ADC systems, one must take into account the "degree of difficulty" of the process technology and layers to which it is applied. Second, the accuracy data obtained is based on from 14-25 lots (depending on the layer), providing a closer "production estimate" of the ADC system accuracy. In spite of the degree of difficulty, the overall accuracy (i.e., agreement between the defect expert and the ADC system) ranged between 70% and 90% on all layers (Fig. 4). In addition, the relative accuracies of individual defect types were all within a similar range; good accuracy in one defect type and poor accuracy in another would generally imply inadequate ADC performance.
We performed an internal study at Motorola to assess the consistency of human classification. Multiple review technicians were asked to classify the same wafer. The results showed an accuracy/consistency (relative to the defect expert) of anywhere between 30% and 70%. The accuracy of the KLA ADC system was therefore judged to be superior to that of human operators. As a result, we have been able to eliminate most manual classification at our APRDL fab. The ADC system performance is verified approximately once a month, on a sample of wafers, to ensure that the machine performance has not drifted for any reason. Stored images are simply reviewed and manually classified, so verification does not require any stoppage of lots in the line.
Equal front- and back-end performance
Equal front- and back-end performance is also important for an ADC system. If the accuracy on different layers varied greatly, or was less consistent than manual methods, we would be unable to entirely eliminate manual classification. Thus, the nearly identical performance on the Active and Metal 2 layers (Fig. 4) is especially noteworthy. Active is a simple, single-layer, front-end step relative to the more complicated, multilayer Metal 2 step. Since these wafers were processed by advanced 0.25-?m CMOS technology, additional complications such as color variation resulting from the chemical mechanical polish (CMP) layers were introduced. For Active, this noise results in a lighter or darker overall background among the comparison die that could lead to significant gray-level differences. For Metal, the gray-level appearance of the previous layer(s) visible through the interlevel dielectric varies widely from one die to the next. In color images, these differences are represented by different colors of the films and layers. In addition, the Metal 2 layer includes noise from the metal grain structure at both the current and the previous layer of metal.
For automatic classification, the defect image and a reference image from an adjacent good die must be captured and compared. The two images are aligned and subtracted. The resulting difference image is then broken down into its constituent features, and these features are analyzed to produce a defect classification.
With complicated CMP layers, much of the noise such as color variation is redetected as part of the defect image. One would expect redetection of the actual defect (and therefore, ADC) to fail. This problem was solved using advanced brightfield image processing algorithms developed for and integrated into the 2132 inspection tool. These same algorithms were then sent to the ADC tool. So, the defects were successfully redetected, in spite of image differences introduced by CMP and grain.
Total time to results
 |
Figure 5. Total time to results, with and without ADC.
The total time to results, defined here as the time taken to acquire data that an engineer can use to start appropriate defect reduction actions, showed a 3? improvement. For manual or automatic classification on a review station, the total time to results is the sum of:
the inspection time;
the queue time between the inspection tool and the review station;
the time required to load and align the wafers on the review station; and
the time required to perform manual or automatic classification.
In the case of a KLA 2132, using manual classification for three wafers in one lot, and assuming 100 defects were classified/wafer, the total time to results was 127 min. To perform the same task using the integrated ADC tool, the total time was only 68 min. Using ADC with the KLA 2135 inspection tool, which has a higher wafer throughput, reduced this number to 38 min (Fig. 5).
Defect baselining
Defect baselining is an additional benefit of integrating ADC into the defect inspector. In the past, defects were typically reviewed only in those cases in which the lot defect count was out of control, or for engineering analysis experiments. Defect review was limited by the significant overhead and extra handling incurred in lot throughput (i.e., when reviewing wafers on an off-line manual review station). Since the integrated ADC system performs automatic classification at 2 sec/defect, data may be gathered for persistent problems, as well as excursions.
 |
Figure 6. Baseline performance for killer defect types, ranked by layer and type.
Baselining, which has allowed us to rank persistent killer defects by type and layer (Fig. 6), is valuable for any defect line monitor strategy, and can be especially helpful in older, high-volume fabs that are running more mature technologies.
Conclusion
The KLA ADC system provides greater accuracy and consistency than previously possible with human operators. It also shows that all layers can be automatically classified. A speed improvement of 3? in the entire inspect/classify step is achieved.
The integration of inspection, classification, and interface to an analysis system into a single tool has produced a process-monitoring system capable of keeping up with the output of today`s high-speed/high-sensitivity defect inspection systems. This integration has been simplified because all three components (detection, ADC, and analysis) have been managed by the same vendor. A high degree of accountability and motivation is important to ensure all components work together harmoniously. In the future, the need for human operators in this time-consuming, subjective, and difficult task may be all but eliminated.n
Acknowledgment
We would like to acknowledge the work of the interns supplied by KLA Instruments, Rohit Chawla, Sheethong Ho, and Chung Fu, and our own technician, No Nguyen, whose efforts were instrumental in bringing this system into production application. We would also like to acknowledge the direction and support of our management team, Fabio Pintchovski and Lou Parrillo.
LOUIS BREAUX received his BS degree in physics from the University of Notre Dame in 1982, and his master`s degree in physics and PhD degree in electrical engineering from the University of Texas at Austin in 1985 and 1989, respectively. He is the primary yield enhancement engineer directing the in-line monitoring program for Motorola`s APRDL. For the past three years, he has been responsible for in-line
monitoring and end-of-line electrical failure correlation, and has been working on an automated defect classification solution for in-line monitoring. Motorola, 3501 Ed Bluestein Blvd., MS K-10, Austin, TX; ph 512/933-7956, fax 512/933-5262, e-mail [email protected].
DAVID KOLAR received his BS degree in chemistry from Miami University in 1977, and his MBA degree from Arizona State University in 1986. He is the yield enhancement manager for Motorola`s APRDL, and has been working on an automated defect classification solution for in-line monitoring since late 1993.