Improving wafer yields at low K1 with advanced photomask defect detection
06/01/1998
Improving wafer yields at low k1with advanced photomask defect detection
Anthony Vacca, KLA-Tencor Corp., Santa Clara, California
Benjamin Eynon, DuPont Photomasks Inc., Reticle Technology Center, Round Rock, Texas
Steve Yeomans, Micron Technology Inc., Boise, Idaho
When the optical resolution is comparable to the illumination wavelength, small variations in photomask dimensions cause disproportionately large changes in wafer CDs. Photomask CD errors =75 nm can print as repeating defects on wafers. This paper describes an investigation of mysterious "stain" defects, which led to the development of a new, advanced line-measurement algorithm (ALM) capable of detecting errors in photomask CDs at the 75-nm level.
To increase the performance potential and life span of current lithography equipment, semiconductor manufacturers have been optimizing lithography by using resolution enhancement techniques such as off-axis illumination, employing higher-contrast photoresist, and utilizing photomasks with optical proximity correction and phase shift. Because the reduction in feature size has recently outpaced exposure wavelength reduction and numerical aperture (NA) increase, semiconductors are being manufactured with optical images [1] having an ever decreasing k1 factor [k1 = (CD)(NA)/l]. Previous work has shown that low-k1 imaging conditions (k1<0.7) produce a nonlinear relationship between changes in photomask and wafer CDs [2]. This "mask error enhancement factor" (MEEF) converts previously unimportant photomask fluctuations into repeating defects.
There is currently a disparity between the capability to detect photomask CD errors (currently ?150 nm) and the detection capability necessary to ensure proper wafer functionality at low k1. While high-end photomask manufacturers are capable of meeting a total CD uniformity specification of approximately 40 nm - as determined by sampling <1000 points with a metrology tool - this methodology will only detect a localized CD error if it occurs at the precise point being measured. In contrast, a total pattern inspection system views billions of points and can ensure detection of all localized errors within the detection and review capability of the system - typically ~150 nm. The newly encountered MEEF effect requires metrology-tool accuracy for all small patterns on the reticle.
 |
Figure 1. Wafer bridging caused by reticle error.
Wafer CD errors
The need for improved photomask inspection capabilities became apparent recently when localized CD errors began printing repeatedly over 5-2000 ?m wafer areas during the production of high-end memory devices. In some cases, the errors were large enough to cause feature bridging (Figure 1). Engineers from DuPont Photomasks, KLA-Tencor, and Micron Technology collaborated to identify the mysterious source of these defects, which initially appeared during automated wafer inspection. Because the errors repeated in the same areas of the affected die, there was reason to believe they originated on the photomask. No photomask defects had been discovered during inspections performed at either the photomask supplier or the wafer fab quality assurance operation. However, due to the nonlinear relationship between photomask and wafer CDs at low-k1 factors, it seemed possible that the photomask defects were sufficiently small to elude detection by conventional means.
"Stain" defects?
The wafer coordinates of the defects were transposed to photomask coordinates for review under a microscope. Figure 2 shows the resulting large regions of dark blotches that were visible in a microscope with a 2.5? objective under brightfield illumination. At most (but not all) defect coordinates, some type of "stain" was visible. Because these anomalies were printing on wafers but were not detected by photomask inspection equipment, the collaborators launched an investigation to determine the nature of the defects. Were they the result of contamination, mask substrate defects, pattern generation errors, processing errors, or something else?
 |
Figure 2. Visible under low-power magnification is a "stain defect" that was due to localized reticle errors.
Since visually the problem appeared to be a "stain," an inspection was ordered using a prototype KLA-Tencor STARlight 300HR system, which analyzed both transmitted and reflected light at high resolution. The inspection failed to locate any contamination defects within the "stained" areas using the TR101 algorithm and a 0.25-?m inspection pixel, the highest magnification then available.
 |
Figure 3. a) Symmetrical and b) asymmetrical alignments of optical (gray) and reference (blue) images used in detection method A.
The STARlight image acquisition system was then used to capture gray-scale images for a manual inspection of the transmission and reflection properties. Again, no defects were apparent in the "stained" areas. The problem was not caused by contamination or substrate defects. The focus of the investigation then shifted to feature-size and pattern-shape anomalies.
CD Errors
The captured gray-scale images were next processed through an off-line defect algorithm simulator that allowed for both automatic and manual image comparison. As expected, the automatic option detected no errors. However, manual comparison revealed slight differences in CD linewidth between the two images. The differences were verified by capturing and comparing numerous images in both the defective and nondefective areas. CD measurements taken with a KMS 310 RT confocal mask metrology system confirmed the suspicion that the defects were not actually stains, but rather localized CD errors with magnitudes ranging from approximately 15 to 150 nm.
What appeared to be blotches in Fig. 2 were actually localized CD errors of 100-125 nm. The defective regions appeared as blotches under low magnification, because so many mask features were off-size that the average transmission and reflection of the entire area changed. Such a change could easily be seen in the regular pattern of a memory array, but might not be easily detected on a logic chip pattern. Of course, isolated CD errors of the same magnitude would not affect a low-magnification image, which explained why some of the repeating defects were not visible with a low-power microscope. Once the project identified the defects as pattern-size errors, the focus shifted from investigating to devising automated techniques to detect these anomalies in a production environment.
Algorithm development
KLA-Tencor engineers carefully studied the performance limitations of the current algorithms, designed to detect =200-nm CD errors, along with the characteristics of the localized CD defects. Most photomask pattern inspection tools find defects by comparing digitized images of one die to the next (die-to-die inspection) or a photomask image to the design data (die-to-database inspection). When the difference between the images is above a threshold, the inspection tool reports a defect. In order to find very small defects, the thresholds must be set as low as possible. At very low thresholds, inherent image noise sources (such as optical distortion, mechanical vibration, pixel quantization, misalignment, etc.) trigger false defect events. Therefore, in order to avoid false detections, the threshold must be set above the sum of all noise sources. The only way to lower the threshold without triggering false detections is to lower the sum of the noise sources.
 |
Figure 4. The ALM100 defect-detection method that measures differences in CDs between a reference image and an optically captured image.
 |
Figure 5. A localized (~100-nm) CD error detected as 25 defective pixels by the ALM100 algorithm on a production photomask. The image is a transmission micrograph with a red box indicating the defect location.
Figure 3 shows a hypothetical situation where one feature of a photomask is 80 nm smaller than the design database size of 800 nm, and the two features are overlaid. One way (method A) of detecting this type of CD error is to mark the location of every edge within a die and compare to the nearest edge location within the reference die (or database). In Figure 3a, the photomask image is overlaid symmetrically with the design data for comparison. While the design size is 800 nm and the photomask image is 720 nm, resulting in a CD error of 80 nm, method A reports two separate edge placement errors of 40 nm under symmetrical alignment conditions. Given a defect threshold of 75 nm, no defects would be reported. In the case of asymmetrical alignment (see Figure 3b), the same images will produce one edge placement error of 0 nm and a second placement error of 80 nm resulting in detection of the defect. Hence, the CD-defect sensitivity of method A is sensitive to image alignment.
The new ALM100 method, developed as a result of this investigation, measures die CDs and compares those measurements to reference die (or database) CDs in a way that eliminates alignment sensitivity. The same images analyzed by ALM100 produce one measurement for every two edges (Figure 4). Therefore, regardless of image alignment, the measured CD difference (which is compared to the 75-nm defect threshold) is always 80 nm, resulting in 100% defect detection. This change in measurement methodology, combined with proprietary CD-measurement-noise reduction, differentiates ALM100 from previous defect detection strategies.
CD sensitivity
To verify the accuracy of ALM100, the photomasks that were found to be defective were re-inspected. In one case, every defective region identified by the wafer fab was transposed to photomask coordinates and compared to defect coordinates produced by the ALM100 at its most sensitive setting. The ALM100 algorithm found all of the defect locations identified by the wafer fab with the exception of one defect later identified on the KMS 310 RT metrology tool to be a 15-nm total CD error. One of the defects captured is shown in Figure 5, which is actually a portion of the KLA-351 review screen. The defective area (which contains 25 pixels above the defect detection threshold) is encompassed by a large red defect-bounding-box. The operator also has the option to view the locations of the defective pixels, which helps in proper disposition.
 |
Figure 6. The four patterns of The_Owl test mask A-D.
Verimask 890EX
Many reticle inspection tools are calibrated with programmed defect test masks made by DuPont Photomasks: the Verimask series. Rows Q and R of Verimask 890EX contain programmed CD errors from 80-800 nm on isolated 800-nm lines. These rows were inspected in die-to-die mode 10 times using the ALM100 algorithm set to the minimum CD error tolerance. The ALM100 produced a 100% detection rate of programmed =80-nm defects as computed by the KLA90 sensitivity analysis tool. Since all defects were detected, the minimum CD error detection could not be determined. It became apparent that a new test vehicle was needed to quantify and calibrate equipment sensitivity to such small photomask CD errors accurately.
 |
Figure 7. Wafer CD vs. reticle CD for pattern A on The_Owl test mask. The increased slope of the measured data (blue with points shown) compared to the 5:1 demagnified CD line (green) documents the MEEF. High- and low-CD uniformity limits (?25 nm) are shown by the straight red lines.
The_Owl Test Mask
The_Owl test mask contains 1.75-?m geometries in 100 ? 100-?m cells arranged as an error cell positioned between two reference cells and separated by 10 ?m. The three-cell arrangement was repeated 15 times along the x-axis with error geometry sizes varying from -175 to +175 nm in 25-nm increments from the nominal baseline. Figure 6 shows the four different geometry types that were replicated along the y-direction using this scheme.
To verify that the designed CD errors were processed correctly, the KMS 310 RT measured the vertical lines in Fig. 6a. Differences between the designed and measured "errors" on the test mask were consistently <7 nm. The mask was then printed on a wafer using an i-line stepper under the conditions shown in Table 1.
 |
The developed, etched, and resist-stripped wafer was then measured using a CD scanning electron microscope (SEM). Figure 7 compares the resulting wafer CD error with the photomask CD error based on pattern A in The_Owl mask at a nominal CD of 290 nm (k1 = 0.43). The chart also graphically illustrates the nonlinear relationship (MEEF) between wafer and photomask CD errors.
Assuming a CD specification of ?25 nm on the wafer, the linear 5:1 reduction ratio predicts that deviations of up to ?125 nm would be acceptable on the photomask. However, Figure 7 shows that 25-nm and 50-nm photomask errors produce corresponding 20-nm and 28-nm wafer errors. Therefore, pushing stepper resolution limits can cause very slight photomask CD errors to affect wafer CDs much more dramatically than would be predicted by the geometrical demagnification of the stepper. Indeed, assigning all the wafer CD error budget to the photomask would still result in a mask CD specification of ?35 nm! Success in achieving high wafer yield under these extremely low k1 conditions requires tightening the CD error specifications and fully qualifying the photomask.
Table 2 shows the defect capture rate for the ALM100 algorithm at the most sensitive setting using The Owl test mask. The gray shaded areas indicate 100% capture, where any size CD error detected in a cell counted as detection of the region.
 |
Production environments
Although Table 2 shows 100% detection of all defects ?50 nm or larger, the 25- and 50-nm detections are not considered to be of practical sensitivity in a production environment because those regions are typically marked defective by only a few pixels. In production mode, an equipment operator would have difficulty separating true defects at that level from false detections. In contrast, the =75-nm defects are marked with a sufficient number of defective pixels to enable correct defect disposition.
Currently available automated photomask inspection equipment is not capable of reliably detecting CD errors below 150 nm. The improved sensitivity provided by the ALM100 algorithm will minimize wafer yield loss due to localized photomask CD errors. KLA-Tencor currently plans to make product versions of the algorithm and test mask available during this calendar year.
Conclusion
The industry-wide practice of employing advanced resolution techniques to extend the production capabilities of today`s wafer steppers has resulted in a nonlinear relationship between photomask and wafer CDs. Consequently, small CD errors on photomasks that were previously innocuous are now printing on wafers, resulting in lower production yields and/or reduced device speed performance. These photomask CD errors are too small to be detected by currently available inspection equipment.
Close collaboration between an inspection equipment supplier, photomask producer, and semiconductor manufacturer has resulted in the development of a new algorithm, the ALM100. This algorithm enables detection of minute photomask CD errors that have caused wafer-pattern bridging in manufacturing. The companies` cooperative approach illustrates the need for close working relationships between suppliers and customers in the ongoing development of deep-submicron manufacturing processes.
Acknowledgments
The authors thank Robert Terhune, DuPont Photomasks; Steve Buchholz, Bill Broadbent, James Wiley, KLA-Tencor; and John Arnaud, DPI Reticle Technology Center. The STARlight 300HR system is a trademark of KLA-Tencor. The Verimask series is a trademark of DuPont Photomasks.
References
1. A.K. Wong et al., "Lithographic Effects of Mask Critical Dimension Error," Proceedings of SPIE, Vol. 3334, paper 10 (to be published in 1998).
2. P-y. Yan, B. Hainsey, J. Farnsworth, J. Neff, "Submicron Low-k1 Imaging Characteristics Using a DUV Printing Tool and Binary Masks," in Optical/Laser Microlithography VIII, ed., T. A. Brunner, Proc. of SPIE 2440, p. 270, 1995.
Anthony Vacca received his BS degree in electronics engineering technology from the DeVry Institute of Technology in 1985. He then joined KLA Instruments and has held numerous positions including design, manufacturing, and applications engineering. He is the technical director for the DPI strategic business unit of KLA-Tencor. KLA-Tencor Corp., 1701 Directors Blvd., Suite 1000, Austin, TX 78744; ph 512/462-6354, fax 512/259-0285, e-mail: [email protected].
Benjamin Eynon received his BS degree in microelectronic engineering from the Rochester Institute of Technology in 1987. He has worked as a process engineer, technical marketing engineer, and manufacturing manager for semiconductor and photomask operations. Eynon is a Research & Development Group leader for DuPont Photomasks Inc.`s Reticle Technology Center and is based in Round Rock, TX.
Steve Yeomans received his BS degree in chemical engineering from Rice University in 1983. He joined Micron Technology in 1984 to work in the Photo Engineering Department of Fab 1, where he was involved in wafer photolithography process control and optimization. In 1990, Yeomans transferred to the Mask Engineering Department at Micron. He is responsible for mask vendor liaison and for defect control and reduction on 5? stepper reticles.