Single-pass die-to-database tritone reticle inspection
05/01/2007
Tritone reticle designs present many challenges for both photomask manufacturers and defect inspection equipment suppliers. From a fabrication standpoint, multi-write and process steps for tritone layers add levels of complexity and increased costs not encountered with most traditional binary (two-tone) masks. For inspection tools, the presence of three distinctive light levels presents a challenge for algorithms originally designed to inspect gray scale data between two tones (black and white), especially in the case of database transmitted light modes. In recent years, the mask industry has used various workarounds to address these challenges, but these have certain limitations in that reticle throughput is either reduced or sensitivity is compromised.
Motivation for study
While most die-to-die inspections on tritone reticles produce successful results using binary algorithms, database inspections typically require two separate recipes to reveal all lithographically significant defects. Using conventional binary database patterns to inspect tritone reticle images presents an obvious data discrepancy. Unlike a die-to-die tritone inspection, in which both reference and test images each contain three distinct light levels, database inspections with reference data containing only two tones produce too many false defects. This outcome is especially true when inspecting 193nm embedded attenuated phase shift masks (EPSM) illuminated by 365nm or 257nm laser light sources. As a result, many mask suppliers use two main workarounds to circumvent this challenge.
Do not inspect region (DNIR). The first workaround involves placing a DNIR around any chrome feature present in the inspection area (IA). By eliminating chrome patterns, the third tone is removed from the optical image, allowing conventional binary database algorithms to be used. While this technique does eliminate the extra tone, the tradeoff with this approach is two-fold.
The first and most obvious concern is that any features within a DNIR border are not inspected. While most chrome structures of a simple tritone reticle appear only in the scribe and border locations, some chrome patterns may be part of the primary design and must be inspected to ensure that they contain no defects.
The other tradeoff is cycle time. The process of adding DNIRs to a recipe is very time consuming. The extra steps needed to implement DNIR locations during recipe creation lowers productivity. While some DNIR borders can be defined using offline database images, it is common practice to identify DNIR locations while the reticle is loaded on the inspection station. This method is often preferred because the live camera image (in transmitted light mode) clearly reveals the dark chrome patterns to include within the DNIR. However, this approach takes away valuable inspection time from the reticle inspection tool, further decreasing overall productivity.
Modified light calibration (MLC). The second workaround involves lowering the transmitted light level of the shifter material using inspection station software to create a “pseudo-binary” optical image. As indicated in Fig. 1a, shifter material normally produces a “gray” transmitted light level not present in conventional binary (two-tone) database images. Forcing the shifter light level to a lower (darker) value, as indicated in Fig. 1b, creates a pseudo-binary image compatible with dual-tone binary database patterns. The result of this MLC technique is that both shifter and chrome features appear dark when viewed with transmitted light. As such, no DNIRs are needed since both shifter and chrome features have nearly the same dark tone compatible with binary database patterns.
 Figure 1. Transmitted light profiles for a) standard light calibration and tritone image, b) modified light calibration and “pseudo-binary” image. |
The main disadvantage of this approach is that there is a loss of sensitivity at the chrome and shifter interface. The lack of dynamic range between these two layers reduces image contrast to a point where defects, such as residual chrome on the shifter surface, are not detected. Other defect types not found using this workaround include residual chrome on shifter/quartz edges and chrome pinholes that expose shifter material.
Whether using the DNIR or MLC approach, there is an inevitable tradeoff in either throughput or sensitivity that must be addressed as part of any thorough inspection program. These limitations associated with using a binary database workaround to inspect tritone reticles require an additional inspection for either the DNIR areas or the shifter/chrome regions, creating an overall dual-pass strategy. It is because of these limitations with either workaround that a single-pass tritone database solution was evaluated with the findings posted here.
Theory of operation
To implement the tritone database capability on KLA-Tencor’s TeraScan (5XX Series) reticle inspection system, two important changes were made. First, database images needed to differentiate the individual tones associated with first and second write level data preparation (data prep) processes. The second fundamental change was with the release of a new database algorithm that could process tritone pattern data to perform alignment, image calibration, and defect detection operations.
 Figure 2. Tritone reticle transmitted light images for a) a binary database, b) tritone reticle optical, c) tritone database. |
Figure 2a depicts a conventional binary database image. Compared to a tritone reticle optical image (Test) in Fig. 2b, the tone mismatch described in the previous section becomes quite apparent. However, when light levels for both first and second level write patterns are maintained, a database tritone (dbTt) reference image can be created, as shown in Fig. 2c. This exact match between the dbTt (where “T” stands for transmitted and “t” for tritone) reference data and test image eliminates the need for either DNIR or MLC workarounds.
Creating tritone reference database images involves a two-stage data prep process. First, both 1st and 2nd level pattern files are prepped individually and then merged to produce one tritone database pattern file using the KLA-Tencor TeraPrep system. The end result is a reference database image containing three separate tones that can be directly compared to a tritone optical image.
In addition to changes in the data prep system, a new tritone database inspection algorithm was developed. With this algorithm, a single-pass database inspection strategy for tritone reticles can be achieved. The dbTt algorithm performs bias and rounding routines similar in concept to the image calibration capability found on previous generation binary database algorithms to adjust the reference image for process-induced pattern changes.
Three new detector categories are introduced in the UCFdbTt45 algorithm to accommodate defect detection among three tones. The Hi-Res1 (pattern) detector includes the following new groups:
- Clear chrome edge: Cr/Qz edge detector
- Chrome halftone edge: Cr/MoSi edge detector
- Triple point: detector for regions where Cr/MoSi/Qz patterns appear in the vicinity of each other
As with other algorithms, defects found during a dbTt inspection are sorted into various bin categories for defect review. Similar to other 5XX detector sliders, a threshold setting of 100 provides maximum sensitivity while a value of zero effectively disables that detector.
Experimental setup
Evaluating the capability of the dbTt inspection mode was divided into two main sections: 1) test reticle performance and 2) production reticle performance. From these two study groups, both sensitivity and throughput measurements were obtained, providing initial information regarding the performance of the dbTt mode.
Test reticle (sensitivity comparison). The SPICA200V6.2 is a test mask used to validate the sensitivity and false defect performance for the 150nm, 125nm, and 90nm views for 5XX pattern inspection modes. Since this 193nm EPSM (6%) tritone mask contains both shifter/quartz and chrome/shifter/quartz programmed defect arrays, capture rate capabilities for both binary and tritone database inspection modes could be compared. In this article, P125 and P90 inspection view performance was evaluated using both binary and tritone database modes.
Ideally, both binary and tritone database inspection modes should have the same defect capture rate. Theoretically, however, it is understood that there would be a slight loss of sensitivity with the dbTt mode. This phenomenon is due to the reduced dynamic range (less image contrast) inherent with any tritone inspection mode (Fig. 1a). For both binary and tritone database modes, the P125 inspection view (125nm pixel size) was used to inspect the 320nm region of the SPICA200 mask while P90 (90nm pixel size) performance data was obtained from the 260nm test structures.
 Figure 3. Throughput comparison overview. |
Production reticle (throughput comparison). For production reticle evaluation, two simple (no chrome in the active area) tritone masks were selected to measure the throughput advantage of single-pass dbTt performance as compared to results from both dual-pass binary database inspection strategies. As shown in Fig. 3, the first case study used a 65nm node tritone reticle to obtain single-pass dbTt results to compare against dual-pass results using the MLC workaround for the P90 view. The second test included a 90nm tritone reticle to measure single-pass dbTt performance against dual-pass inspection efforts with the DNIR workaround for P125.
For all throughput measurements, the TeraScan was configured as a T4 system, which indicates a specific image computer configuration. IA sizes were ~100×115mm for all inspection modes.
Test reticle results: sensitivity comparison
P125 results. The 320nm region of the SPICA200 was inspected 10 times in the EPSM programmed defect array using the UCFdbT45 binary database algorithm. These results produced an average defect count of 340. Using the UCFdbTt45 tritone database algorithm, the same “EPSM only” region was again inspected 10 times, generating an average defect count of 333 detections. As anticipated, there was a slight loss of sensitivity (seven counts on average) with the dbTt inspection mode. Both sets of data were obtained using maximum sensitivity settings for the Hi-Res1 and Hi-Res2 detectors.
Defects found by the dbT-binary mode that were not captured by dbTt included fewer pinholes and edge defects located within the smallest extents of the programmed defect grid. The “gap” in pinhole detection for both results is typical with any transmitted light inspection algorithm and is commonly addressed using alternate inspection modes [1]. No false defects were observed in any of these P125 runs.
Additional dbTt results were obtained with the “EPSM Only” IA extended to include “chrome on shifter” programmed defects. Using 10 sequential inspections, an average defect count of 400 was produced, also using the maximum sensitivity settings for the Hi-Res1 and Hi-Res2 detectors. This total includes the 333 “EPSM” detections in addition to 67 chrome-on-shifter programmed defects. These types of chrome-on-shifter defects are not consistently found when using the MLC workaround to inspection tritone reticle using a binary database.
P90 results. The 260nm region of the SPICA200 was inspected 10 times using the UCFdbT45 binary database algorithm resulting in an average defect count of 347 detections. Using the UCFdbTt45 algorithm and tritone database, the same inspection area was inspected 10 times and produced an average defect count of 337 detections. As anticipated, there is a slight loss of sensitivity (10 defects on average) using the dbTt inspection mode. Both “EPSM Only” data was obtained using maximum sensitivity settings for the Hi-Res1 and Hi-Res2 detectors.
The dbTt results were also performed with the “EPSM Only” IA extended to include “chrome on shifter” programmed defects. These 10 sequential inspections produced an average defect count of 407 also using the maximum sensitivity settings for the Hi-Res1 and Hi-Res2 detectors. This total includes 337 “EPSM” detections in addition to 70 chrome-on-shifter programmed defects. No false defects were observed in any of the P90 runs.
Some of the smallest programmed defects found by the dbT-binary mode were not detected by dbTt due to the reduced dynamic range. Although no thorough investigation was performed, these defect types are not lithographically significant for most wafer fab requirements. Most binary database inspections typically require slightly reduced sensitivity settings on production reticles to intentionally ignore noncritical defects similar in size and location to those presented here.
Production reticle results: throughput comparison
Case Study 1 - P90 dbTt vs. db-binary MLC results. The inspection comparison results for this reticle are provided in Fig. 4 with values rounded to the nearest whole minute. The total throughput needed for the dual-pass inspection, as seen in the top row of the chart, was 395 min. By comparing this data against the single-pass dbTt results of 230 min, the throughput advantage of the dbTt inspection mode can be seen. Data from the various inspection steps are contained within the table in Fig. 4 and were used to generate this chart.
 Figure 4. P90 results (dbTt vs. db-binary MLC). |
The “Recipe 1” results reveal dbTt taking 10 min longer to create for the initial setup period (16 min vs. 6 min). This is due to the extra steps necessary to produce the merged tritone database file. The “Cal 1” segment includes plate alignment, light calibration, and image calibration routines, which were virtually the same for both modes (2 min difference). Total inspection time “Insp 1” for both modes was also similar (within 10 min of each other).
The “Recipe 2” category was 0 min for the dbTt since a second inspection was not required using this single-pass strategy, which is also the case for the “Cal 2” and “Insp 2” entries. By using the dbTt single-pass solution, a second inspection is not required, which is the primary advantage of this inspection mode. The second set of data for the db-MLC inspection strategy was obtained from a die-to-die transmitted light inspection (161 min) needed to capture chrome or contamination defects on MoSi not found using the MLC workaround.
During this experiment, it was observed that the defect review period was essentially the same for both single- and dual-pass strategies and therefore these results are not included. It should be noted, however, that additional time and resources are needed to manage two IRs associated with the dual-pass inspection strategy for a single tritone reticle. Mask manufacturers would need to pay special attention to ensure that both sets of IR coordinates are provided to defect repair tools, since this is normally a single-step procedure.
Another potential glitch with using the dual-pass inspection strategy for tritone reticles is with production floor management software normally designed to interface with one inspection output file such as an AutoPrint IR text file. To ensure that two IRs are tracked by such software may require manual intervention or programming modifications to such a software system. The extra time needed to administer these post-inspection steps was relatively difficult to measure precisely, but it should be considered in a dual-pass inspection strategy.
An item worth mentioning regarding the creation of merged dbTt pattern files has to do with pattern polarity of the first and second level write layers. Attention must be given to the selection of correct polarity (positive or negative) during data prep to match the lithography process used on the reticle (positive or negative resist). If either layer contains the wrong polarity for the type of resist used, a reverse tone image can be generated in the dbTt merge file, causing problems during image calibration or excessive defects during inspection.
MLC: chrome on shifter sensitivity loss
Figure 5b reveals a sensitivity problem when using the MLC workaround. The non-shaded region (see arrow) reveals a capture rate difference for chrome-on-EPSM (shifter) programmed defects within columns R, S, and T. When compared to the dbTt results in Fig. 5a, it becomes apparent that the db-binary inspection mode (involving the MLC workaround) does not provide a complete inspection solution for inspecting tritone reticles. The results in Fig. 5 clearly illustrate the need for a second inspection (hence the dual-pass strategy) when attempting to inspect tritone reticles with the db-binary inspection using the MLC workaround.
 Figure 5. Chrome-on-EPSM capture rate comparison for a) dbTt mode, and b) db-binary (modified light calibration) mode. |
The low contrast between chrome and shifter materials (see Fig. 1b on p. 56) produced from this workaround creates unreliable sensitivity performance for chrome-on-shifter program defects. No defects in column S of the test reticle (chrome spots on EPSM) were found using the db-binary mode with the MLC workaround.
Furthermore, chrome-on-shifter edge test reticle defects could not easily be seen because both chrome and shifter materials are reduced to nearly the same black level when using the db-binary mode with the MLC workaround. The same imaging challenge also impacts pinhole-in-chrome (with exposed shifter) defects.
Case Study 2 - P125dbTt vs. db-binary Results (DNIR). The total throughput needed for the dual-pass inspection was 279 min. The throughput advantage of the dbTt inspection mode can be seen by comparing this data against the single-pass dbTt results of 144 min.
The “Recipe 1” results revealed dbTt taking 16 min less to create for the initial setup period (31 min vs. 15 min). This is due to the extra steps needed to define the DNIR portion of the recipe. The “Cal 1” segment includes plate alignment, light calibration, and image calibration routines which were virtually the same for both modes (1 min difference). Total inspection time “Insp 1” for both modes was similar as well (within 9 min of each other).
The “Recipe 2” category was 0 min for the dbTt since a second inspection was not required using this single-pass strategy, which is also the case for the “Cal 2” and “Insp 2” entries. By using the dbTt single-pass solution, a second inspection is not required, which is the primary advantage of this inspection mode. The second set of data for the db-DNIR inspection strategy was obtained using a transmitted light db-binary inspection (89 min), needed to augment the dual-pass inspection strategy for this tritone reticle (which inspected the border and scribe areas with DNIRs for the main pattern die).
It was observed that both data prep and defect review segments were essentially the same for both single- and dual-pass strategies; therefore, these results are not included. Again, note that additional time and resources are needed to manage two IRs associated with the dual-pass inspection strategy for a single tritone reticle. Like in the previous case study, some sort of manual intervention or programming modifications may need to be factored into the process.
Conclusion
This article examines the benefits of using a tritone database inspection algorithm from both productivity and sensitivity standpoints as compared to results obtained from using standard workarounds and existing binary inspection modes. The results and conclusions are based on data obtained from standard test vehicles and several tritone production reticles.
The dbTt mode provided a single-pass solution, offering a throughput advantage over db-binary results on tritone production reticles. As measured in the first case study using the P125 inspection view, the improvement was a factor of 1.93x (144 min for dbTt single-pass compared to 279 min using the dual-pass db-binary MLC approach). For the second case study, the P90 view performance revealed an improvement factor of 1.72× (230 min for dbTt single-pass vs. 395 min for the dual-pass db-binary DNIR strategy).
When comparing single-pass sensitivity performance, dbTt consistently found the chrome-on-shifter programmed defects that db-binary inspections (using the MLC workaround) did not detect when the IA included columns R, S, and T. The defect types found in this region include chrome-on-shifter edge defects as well as residual chrome spots on shifter and chrome pinholes (with exposed EPSM). Any thorough inspection strategy using a binary database approach using either workaround to inspect tritone reticles will require a second inspection to find these types of potentially yield limiting defects.
The dbTt mode was found to be slightly less sensitive to certain defect types when compared to db-binary inspection results on the SPICA200 tritone test reticle. The two categories that reveal a capture rate performance difference are pinhole and on-edge/line-end defects. However, the db-binary mode also did not detect all pinholes present in the 193nm EPSM programmed defect array. Both inspection strategies require an additional inspection strategy, such as a reflected light solution, to eliminate the “pinhole gap” effect inherent with transmitted light inspections on 193nm EPSM masks.
With respect to dbTt performance for on-edge and line-end type defects, the differences occurred with the smallest programmed defect sizes and are not generally considered to be lithographically significant for most wafer fab requirements. Data used to support this claim include edge and line-edge defect measurements ranging in size from 18nm to 23nm. In general, most binary database inspections typically require reduced sensitivity settings for production reticles to intentionally ignore small defects. A detailed printability study of these and other defect types not detected by the dbTt mode, which are commonly found on production reticles, is being considered for future work.
Acknowledgments
TeraScan and TeraPrep are trademarks of KLA-Tencor Corp. This paper is based on a work by B. Reese, J. Heumann, and N. Schmidt, “Single-pass Die-to-database Tritone Reticle Inspection Capability,” appearing in Proc. of SPIE, Vol. 6349, 63493L, 2006, doi: 10.1117/12.686121.
Reference
1. A.D. Vacca, D. Taylor, “Comprehensive Defect Detection Featuring Die-to-database Reflected Light Inspection,” BACUS Symposium on Photomask Technology, 2004.
Jan Heumann received his MS and PhD from the Technical U. of Berlin and is a member of the technical staff at AMTC, Rähnitzer Allee 9, 01109 Dresden, Germany; ph 49/351-4048-264, email: [email protected].
Bryan Reese received his Associates Degree in electronics from DeVry Institute of Technology and his MBA from Hawthorne U. and is a technical specialist at KLA-Tencor in Austin, TX.
Norbert Schmidt received his Diplom-Engineer (FH) title for engineering physics from Fachhochschule München in 1996 and is an applications engineer at KLA-Tencor in Dresden, Germany.