Applying a Defect Spatial Signature Analysis in a Production Fab

As semiconductor devices shrink, the need for tighter process controls becomes more critical. The process engineer must know when problems occur, and be directed to their source. The impact of problems can then immediately be addressed to minimize the effect on other products. Automated defect inspection can generate hundreds of wafer-inspection records per day, each with specific defect patterns.
 
These "defect spatial signature" patterns may indicate a systematic problem with a process or equipment. Because of the sheer volumes of defect data collected in the fab, it is nearly impossible to find signatures among all the data using a manual review process. A viable defect spatial signature analysis (DSSA) system can quickly sift through all the defect data and return consistent classification results. Additionally, the DSSA system can find signatures that humans would miss, and can notify the fab personnel when the signatures occur.

In this article, the signature data from HPL's Sentry DSSA system is compared against a manual signature review process to determine if it would enhance signature capture rate. We discuss the results of this evaluation and the benefits of using DSSA in the fab.

Using manual methods

As part of the normal fab control system, the levels of equipment, chemical mechanical polish (CMP) and handling scratch signatures are monitored in the manufacturing facility on a monthly basis. The manufacturing facility is divided into two fabs, which can be denoted as "fab A" and "fab B": both fabs run the same processes.

Fab A is a recent addition with relatively new processing equipment, compared to fab B. Fab A also uses more wafer-sorting equipment and SMIF containers than fab B. Combined, the two fabs inspect ∼500 wafers per day. The data is partitioned between the two fabs, and the signature levels are reported by each fab separately and in combination.

Before using the DSSA Sentry system, a manual review process was used to detect the signatures. It was impractical to review all 500 inspected wafers per day. Only the maps with defect counts above an established defect density (DD) target were reviewed. This amounted to ∼25% of the inspected wafers. The maps were reviewed using a typical graphical viewer on a GUI display. If the signature was obvious, it would be counted.

Counts for each signature type were plotted on a monthly basis (Figs. 1-5 ). Several months of manual review indicated some differences in signature levels between the fabs, with a tendency toward more scratch signatures in fab B. A difference in signature levels between the two fabs was expected, especially with the equipment and handling scratch signatures. It is possible that the manual review process was biasing the signature counts. It was unclear how much of the signature levels was real or an artifact of the review process.

1. Signature counts for fab A and fab B from manual review of signatures.

2. Signature counts by signature type from manual review of signatures.

3. Equipment signature count from manual review of signatures.

4. CMP signature count from manual review of signatures.

5. Handling scratch signature count from manual review of signatures.

In early 2004, we installed a beta version of HPL's Sentry DSSA system, and then connected it to the defect data management system (DDMS) database. The Sentry system automatically reviewed all defect maps, whether they were above or below the DD targets. The Sentry system is set to ignore maps with <10 defects and >65,000 defects. The entire classification process typically occurs within a few seconds per wafer. Each map could have more than one classification assigned to it if the Sentry system detected more than one signature on the map. The data was stored in the DDMS database (Fig. 6 ).

6. The DSSA Sentry system receives data from the database. The classifications are performed within a few seconds per defect map, and the data is returned back to the database. Classifications may be defined using the Sentry manager. Sentry also supports event-based actions (alarms).

The Sentry system analyzes wafer maps with two distinctly different signature engines. One engine, called wafer signature analysis (WSA), analyzes the global signature pattern for the entire wafer. WSA returns a signature name based on a pattern matched to a template, such as "ring outer," "ring middle" and "center." Users can create or edit templates to define a regional signature. There is no need to create templates for each variation of signature rotation.

The other engine is called cluster signature analysis (CSA). CSA returns data for each recognized cluster, such as shape, length, center of mass, density, and so on. In addition, CSA returns a set of descriptive data for each cluster object, such as "line center vertical," "arc edge," and so on. CSA does not require templates to define the cluster shape signatures.

The data from both engines is used in customizable rules to define the classifications. If the result of the rule is "true," the classification is assigned. If more than one rule is true, multiple classifications may be assigned. Arbitrations between two or more simultaneous classifications are allowed to filter out one or more multiples.

Classifications may also be set using other parameters, such as layer, inspection recipe, inspection tool, etc. Rotation data may be used to set classifications based on specific angles, or the rotation may be ignored to set single classifications for all possible rotation angles. This flexible, highly configurable system is able to support many classification groups with very good accuracy and purity values for each group.

Using DSSA to analyze signatures

Once the early phases of the beta were complete, we set up the Sentry system to automatically classify wafer maps according to the defined classification groups. The system was set to classify 11 individual signature groups. Three classification groups were the focus of this study: CMP scratch, handling scratch and equipment scratch. Examples of these signatures, and their typical accuracies and purities, are shown in Table 1 .

In the normal course of automated classification, some wafer maps were misclassified. This is to be expected for any automated classification system. In the DDMS client, wafer maps run within a specific period of time were retrieved by a signature group. All maps in each group were manually screened to correct the misclassifications and purify the results. Many maps did not require screening, because DSSA had already narrowed the scope of the signature maps significantly. Each DSSA signature that could be validated visually was counted. If a map had more than one valid DSSA signature, it was counted once for each classification group.

It was not always possible to readily see the signatures on the wafer maps during the screening process, even though the Sentry DSSA system indicated signatures existed. The normal reaction was to dismiss these maps as misclassified. However, closer inspection revealed most of these maps had small or faint signatures that the graphical display did not reveal under normal viewing conditions. This display issue occurred on maps regardless of whether they were above or below the DD target levels. Therefore, it was entirely possible that many of the maps that would have been manually reviewed (maps above the DD targets) could have had signatures that were missed by manual inspection.

The effectiveness of the manual review process was now in doubt. The manual capture rate was much lower than originally thought. Based on the number of times closer-than-normal scrutiny revealed signatures on maps that would typically have been missed, the manual capture rate was estimated to be around 20%. This means ∼80% of the maps reviewed manually had signatures that could not be easily seen in the display. By comparing the manual review results to the DSSA results, the effect of the manual bias could be measured. Clearly, the signature results were affected by the inability to sufficiently capture small and faint signatures.

When the DSSA data was added to the signature charts, a dramatic change was seen in the overall signature levels (Fig. 7 ). Unexpectedly, the differences in signature levels between the fabs were less significant with the DSSA data. This may be because of the ability for DSSA to classify all maps instead of only the maps above the DD targets. Of course, a lot of the differences were also caused by the increased capture of smaller, fainter signatures.

7. Total signature count by fab with manual review and DSSA. DSSA implementation led to less difference in total defect count between the fabs.

Looking at the data by signature type (Fig. 8), it can be seen that all levels of signatures increase with DSSA. The equipment signatures increase dramatically (Fig. 9), and handling scratches to a lesser degree (Fig. 10 ). In both of these groups, the ability to detect smaller and fainter signatures on maps that would have lower DD (and would therefore not be manually reviewed) has provided a much clearer idea of the actual levels of signatures in the two fabs.

8. Once DSSA was implemented, it became clear that many more handling equipment defect signatures were detected.

9. The DSSA is able to detect smaller and fainter defect signatures, resulting in higher equipment signature counts.

10. Fab A is newer and has more automated handling than fab B, so it had lower defect density and more defects were detected with the manual process in fab B.

CMP scratches do not appear to have increased as dramatically as the other signatures (Fig. 11 ). This is probably because of the fact that maps with CMP scratches tend to be above the DD levels, and so many of these maps would have been detected during the manual review process before using DSSA. As stated previously, fab A is newer than fab B. It uses newer process equipment and more automated handling equipment. The biggest shift in the signature data with DSSA was with the equipment scratch signatures, and to a lesser degree, the handling scratch signatures. Fab A tended to have smaller and fainter signatures than fab B. This indicates that the control over equipment and handling is better in fab A than fab B. Also, fab B tended to have higher DD levels, so the manual review process found more signatures in fab B. Therefore, the DSSA data provides a more accurate picture of what is really going on in both fabs.

11. CMP signature count did not increase greatly with DSSA, probably because maps with CMP scratches tend to be above the defect density levels, so many of these maps would have been detected during the manual review process before using DSSA.

Benefits of DSSA

Because the DSSA Sentry system can typically classify wafers within a few seconds, the limitation of reviewing only maps that are above the DD targets is no longer an issue. Our analysis revealed the manual review process could only capture signatures ∼20% of the time on the maps above the DD targets. It is clear that the Sentry system provided a much more accurate view of the signature levels in the fab, compared to the manual review process.

To estimate the level of improvement with DSSA over the manual methods, the "effective capture rate" could be calculated for each method using the data generated in this study. The effective capture rates were calculated by multiplying the percent of wafers reviewed by the typical capture rates for each method. The effective capture rate for DSSA was ∼16× higher than the manual review process (Table 2 ).

The difference in effective capture rates implied that DSSA could shorten the time-to-detect (TTD) a potential yield problem (assuming it is signature-related), compared to the manual review process. Because of its speed at classifying wafer maps, the Sentry system could review all maps in the fab in real-time. This makes it an effective monitor for signature-related problems.

Assuming the time-to-fix is held constant, regardless of the signature detection method used, automated DSSA could shorten the TTD by as much as 1 to 3 weeks. If a signature-related problem caused a 5% yield excursion, the shorter TTD could translate into significant savings on a per-incident basis. Of course, these results depend on defect inspection sampling rates and other factors. Even so, the idea that a shorter TTD translates to a cost savings applies to most fabs.

One may argue that clusters, and by extension signatures, typically do not affect many die. So why worry about them? Signatures, no matter how faint or small, may affect yield in two ways:

• They can be an early indication of a problem that may get progressively worse.
• They are an indication of an overall systematic issue in the fab. If they occur frequently, their effect on yield can be felt over many wafers and lots. Small doses with many occurrences can be as bad as, or worse than, large doses with few occurrences.

Conclusion

It is vital for all fabs to immediately know when potential yield problems occur. It is well understood that signatures can be associated with handling, process or equipment problems. DSSA provides better resolution of issues that were previously missed. The ability for Sentry to provide better capture and discrimination of the signatures affords more opportunity to have better control over fab yield.

Based on these results, the Sentry DSSA system will be used as a real-time monitor to detect signature-related problems as they occur in fab A and fab B. The system will feed the MES system to auto-alarm on specific signatures, in addition to the typical DD targets. The signature data will also be used for data mining and correlations to learn the source of the signatures. This learning can be applied for better control over handling procedures, process issues and equipment maintenance.

Acknowledgements

The authors wish to thank the following people for their support in this effort: Gary Green, Ying Mu Zhong, Mark Meister, Liqing Fu, Weiping Gao, Michael Wu, Casey Chou, ZhiJiang Jin, Hui Xing, Arman Sagatelian, Anthony Adamov, Mark Matheney, Rob Simpson and Tom Jacobs.