By Allen Park, KLA-Tencor Corp., Milpitas, CA, United States
EXECUTIVE OVERVIEW A powerful new method can identify systematic defects within a large defect sample, prior to SEM review. By integrating design data with defect data, this method enables accurate binning of randomly distributed structural systematic defects. Instead of relying on inefficient random review sampling to identify defects of interest (DOI), this technique applies a pattern search engine accessing the design files to correlate the DOI to pattern backgrounds, independent of their spatial distribution. Based on this approach we have identified numerous systematic defects including a residue defect. This novel binning technique allows users to quantify systematic defect types quickly and efficiently from wafer maps that consist of random and systematic defects, allowing for prompt corrective action.
Defect inspections performed during process development often result in 105 to 106 defect counts on a single wafer. Such defect data include both systematic and random defects, only some of which may be yield-limiting. The traditional method of reviewing a random sample of only 50–100 defects on the SEM makes it difficult to identify important systematic defects from a defect wafer map.
Systematic defects, generally pattern failures due to process or design marginality or parametric failures due to electrical issues, are growing in importance as a factor in overall yield loss [1]. Pattern-related systematic defects include line-end thinning, necking, CD variations, side profile variation, overlay error, broken lines, and edge residues (Fig. 1).
Figure 1. Sample pattern-related systematic defects. (Images courtesy of UMC)
Such systematic issues can be challenging to identify using in-line defect inspection systems, due to a high volume of other defect types, and noise-related nuisance sources. Since all systematic defects must be identified during process development, inspection recipes are often deliberately set with high sensitivity, even at the cost of potentially including large numbers of nuisance defects. The defect count can be upwards of hundreds of thousands per wafer, especially in the process development environment.
After inspection, defects must be reviewed to identify their type. A typical repeater analysis based on die-to-die comparison may not be sufficient to detect defects such as line-end thinning or broken lines, because the failure sites are not consistent among various die. Such randomly distributed structural systematic failures are compounded when combined with high defect count. Because of the time and effort required, defect review sampling is often limited to 50 to 100 defects per wafer. With 100,000 defects, a review sample of 100 represents only 0.1 % of the total population, enormously diminishing the probability of identifying critical systematic defects during defect review.
To reduce the difficulty of identifying randomly distributed structural systematic failures, a new design-based inspection technique from KLA-Tencor was evaluated. This advanced technique has been used for almost two years, generating valuable results for both 65 and 45nm development. The integration of design data with defect data enabled us to bin defects using the design background as a proxy for SEM review.
Systematic defect: STI residue
Within the shallow trench isolation (STI) process, a residue defect was discovered using random SEM review after inline defect inspection, however quantity and spatial distribution were unknown due to sampling limitation (Fig. 2). The inspection result was first analyzed using a traditional sampling approach. Even with the relatively low total defect counts on the wafer, the review sample of 50 defects/wafer, selected either at random or using defect size information, represented less than 10% of the total population. It was thus very difficult to quantify the occurrence of the residue defect and to identify its spatial signature, and the largest bin in the defect Pareto was the ‘unclassified’ bin. Only a small number of defects were identified as the residue defects in the traditional defect Pareto.
Figure 2. SEM images of the residue defect. (Image courtesy of UMC.)
To understand how the new technique might better quantify the residue defect population, the same data set was analyzed using both defect and design information. Each defect location was associated with, and then grouped by, the background patterns defined in the design. After grouping, a smart sample of defects was chosen for review. With this approach a significantly different Pareto was generated, and not only 50 defects, but all defects in the inspection result were classified. Such a technique provided a unique advantage in selecting the right set of defects for review, optimizing the return on the defect review effort, and quantifying an unknown failure mode that may otherwise have been overlooked.
Figure 3 compares the resulting Paretos using the traditional and new approaches. In the traditional approach, only 5% of the total defects were identified as residue defects, while the majority of defects remained unclassified. By applying the new technique, all defects were classified, and residue defects were identified at >13× the number identified using the traditional approach. With a significantly higher number of defects now identified as the defect type of interest, their spatial signature can be clearly understood (see bottom image in Fig. 3 below).
Figure 3. (top left) Pareto using traditional defect review; (top right) Pareto using new design-based technique; and (bottom) spatial signature of the residue defect (highlighted).
We found that the residue defects were structurally systematic, but spatially random within some of the die. While the defects seemed to occur in a certain pattern within the die, the failures did not occur at the same locations, according to a die-to-die comparison. By applying design data, we were able to identify that certain parts of the design are prone to this type of failure. Figure 4 illustrates the high probability locations of failure sites for the residue defect type.
Figure 4. Illustration of defect using design data. (Image courtesy of UMC)
Conclusion
Using design information associated with the defect locations provides significant advantage in identifying systematic defects. The new technique relies on an inspection tool with sufficient sensitivity and location accuracy. Using the new design-based inspection technique, unknown systematic defects can be identified and quantified quickly, leading to rapid root cause discovery and correction. While the traditional approach typically samples small portion of overall population, using the new approach allows user to sample 100% of defects by using design as proxy to SEM review.
This kind of structurally systematic, but spatially random defect typically occurs within a fraction of the die; because it does not occur at the same locations, a die-to-die comparison is not suitable in identifying the problem. Using the design data as an integral part of the inspection, we significantly increased our ability to identify those parts of the design most prone to this type of failure. ♦
Acknowledgments
The authors would like to thank J.H. Yeh, Hermes Liu of UMC’s CRD YE team, and Dr. Tzou for providing courtesy sample images.
Reference
1. K. Monahan, B. Trafas, “Design and Process-limited Yield at the 65nm Node and Beyond,” SPIE, 2005.
Allen Park received his BS degree in physics from U. of Irvine, California, in 1988 and is now a marketing manager at KLA-Tencor Corp., where he has worked for more than 12 years. Prior to K-T, he worked in process development and yield enhancement at National Semiconductor and Silicon Systems. KLA-Tencor Corp., 1 Technology Drive, Milpitas CA 95035, United States; ph 408/875-5195, e-mail [email protected].