High-Volume Manufacturing Device Overlay Process Control

There is a need to find more comprehensive solutions to characterize and minimize the size and variability of non-zero overlay.

popularity

By Honggoo Leea, Sangjun Hana, Jaeson Wooa, DongYoung Leea, ChangRock Songa, Hoyoung Heob, Irina Brinsterb, DongSub Choic, John C. Robinsonb

aSK Hynix, 2091, Gyeongchung-daero, Bubal-eub, Icheon-si, Gyeonggi-do, 467-701, Korea
bKLA-Tencor Corp., 8834 N. Capital of Texas Hwy, Austin, TX 78759
cKLA-Tencor Korea, Starplaza bldg.., 53 Metapolis-ro, Hwasung City, Gyeonggi-do, Korea

Abstract
Overlay control based on DI metrology of optical targets has been the primary basis for run-to-run process control for many years.  In previous work we described a scenario where optical overlay metrology is performed on metrology targets on a high frequency basis including every lot (or most lots) at DI.  SEM based FI metrology is performed on-device in-die as-etched on an infrequent basis.  Hybrid control schemes of this type have been in use for many process nodes. What is new is the relative size of the NZO as compared to the overlay spec, and the need to find more comprehensive solutions to characterize and control the size and variability of NZO at the 1x nm node: sampling, modeling, temporal frequency and control aspects, as well as trade-offs between SEM throughput and accuracy.

1. Introduction
Advancing technology nodes in semiconductor manufacturing require smaller process margins.  Overlay control is one of the key parameters to control due the tight requirements and strong correlation to yield.  Overlay control based on DI (after develop inspection) metrology of optical targets has been the primary basis for run-to-run (R2R) process control for many years, and has included significant enhancements in precision, accuracy, and throughput.  Scanning electron microscopes (SEMs) have been used at FI (after etch inspection, or final inspection) to provide supplemental overlay data to R2R process control systems on a limited basis due to throughput, cost, lack of material transparency, and inability to rework.   In previous work [1] we described one such scenario where optical overlay metrology is performed on metrology targets on a high frequency basis including every lot (or most lots) at DI.  SEM based FI metrology is performed on-device in-die as-etched on an infrequent basis.  The difference between DI and FI overlay is often referred to as non-zero overlay (NZO), referring to the offset imposed on the DI R2R control loop based on the infrequent FI updates.  Hybrid control schemes of this type have been in use for many process nodes: what is new is the relative size of the NZO as compared to the overlay spec, and the need to find more comprehensive solutions to characterize and control the variability of NZO.

Optical metrology has and continues to be the primary tool for overlay metrology.  It is the semiconductor industry mainstay for high sampling R2R process control, as well as lot, tool, and process disposition.   In most processes, every lot is measured optically at DI.  SEM metrology, on the other hand, has had limited use for overlay.  SEM overlay is primarily used for DI optical accuracy verification and in some cases FI low sampling frequency R2R process control updates to NZO.  Key NZO questions that this investigation addresses include what sampling and modeling are required, how often does NZO need updating, and can we anticipate changes in NZO?


Figure 1: Schematic of automated process corrections update using 5D Analyzer.

Figure 1 shows the processing scheme that was implemented for this investigation.  There are several ways that non-zero overlay process control can be realized.  In some scenarios, the DI raw overlay data is corrected for the NZO bias.  In this case the modeling and APC (automated process control) remain as usual.  In another implementation, the APC system can be adjusted to control to non-zero targets, though this adds complication to the fab systems.  In this scenario, the runtime DI modeled results are compensated, or corrected.  Those corrections are updated on an infrequent basis dictated by the low sampling rate of the FI metrology.  When FI data is received, it is subtracted from the corresponding lot and wafer DI results.  The subtraction can be done via point-by-point subtraction when the DI and FI metrology locations are in close spatial proximity on the wafer, or via model interpolated subtraction in the general case.  The resulting delta needs to be validated for errors, and then combined with historical values via weighted moving average (WMA), exponentially weighted moving average (EWMA), or the like, in order to minimize the impact of noise on the control loop.  Once the weighted delta values are available, the runtime DI control R2R loop can be compensated by the NZO or bias.  This implementation was implemented within KLA-Tencor Corporation’s 5D Analyzer.


Figure 2: Schematic of overlay and NZO budget breakdown analysis using 5D Analyzer.

Figure 2 shows a schematic of the budget breakdown analysis used in this investigation.  An ANOVA (analysis of variance) is used to decompose the total variability in overlay measurements into different variance components.  The analysis can be applied to raw overlay data, but can also be applied to modeled overlay data as well.  In the case where modeled data is used, the inputs are the modeled results evaluated at the measurement locations of the original raw overlay data.  In this situation, the residuals are not included in the analysis, just the modeled results.  While the budget breakdown tool in 5D Analyzer can provide many components of variability, in this investigation we limit our analysis to mean/common, lot-by-lot, and chuck-by-chuck analysis of NZO for the sake of simplicity.

In this study we looked at NZO for three sets of data corresponding to T1x 19 nm DRAM processing.  Two of the data sets, labeled BLC, correspond to double patterning that included full map data.  The third set, labeled SCC, included reduced sampling and SEM OL data only in the Y direction due to the geometry of the layer.  These layer results will be discussed next.

2. BLC Layer Analysis
First we will describe the bit-line contact (BLC) data that were analyzed.  BLC1 and BLC2 comprise 2 separate prints in an interleaved dual patterning scheme.  For each BLC print, we collected 23 lots of data, 1 wafer per scanner chuck, for a total of 46 wafers each.  For this study we collected both X and Y overlay at both DI and FI, over a 5 month time-span.  As indicated in Figure 3, the sampling included full map data, including 2265 points per wafer, and 25 points per scanner field, in a symmetric sampling scheme.  In this case symmetric denotes that all fields have essentially the same sampling.  The DI metrology was performed by Archer™ 500 metrology tools on Archer AIM® optical imaging targets.  The FI data metrology was performed by CD SEM on the device structure.  These data sets provide much more sampling than is typically available for the analysis of NZO, including high wafer and field sampling at both DI and FI, and additionally comprise a relatively long timeframe for analysis of variability.


Figure 3: BLC experimental setup (a) wafer sampling, (b) field sampling, (c) DI metrology with Archer AIM, and (d) FI metrology on the device structure.

Figure 4 includes trend charts for the BLC1 layer, including DI overlay, FI overlay over the 5 month timeframe of this investigation.  Plotted are |mean|+3σ values for X and Y overlay for lots where NZO is to be calculated, which are only a low-frequency small-fraction of the total number of lots being processed.  Variability includes processing effects inside and outside of the litho cell, as well as metrology.  The fab process control seeks to minimize variability, making periodic adjustments to scanner correctables.  On occasion, process tool adjustments are made as part of planned maintenance (PM) and un-planned maintenance.  In Figure 4c, the delta or NZO is plotted over time.  The corresponding DI and FI lots are subtracted, in this case point by point, and the resulting |mean|+3σ values for X and Y are plotted.  Variability is present in the BLC1 data, however, no significant trends or change points are seen in the 5 month time-span.  In the NZO chart, a minor change point is seen, which corresponds to increased sampling of the extreme edges of the wafers.  This change was made in order to better understand the extreme wafer edge behavior, as will be discussed later in this report.  Therefore, other than an assignable change point, no significant trends or un-explained change points were observed.


Figure 4: BLC1 |mean|+3σ trend charts (a) DI overlay, (b) FI overlay, and (c) NZO.

To further understand the trends, or lack of trends in this case, the BLC1 NZO data was modeled with a linear wafer and linear field (W1F1) model, as shown in Figure 5 by lot ID.  Minimal drift is seen, with the small possible exception of intra-field Y rotation (Y_R_Rot) or K6 term having minimal impact.  The data are divided up into groups for further investigation.  Groups 1 and 2 are arbitrarily chosen as the first and second halves of the data before the sampling change point that was discussed above.  The third group is the data after the extreme edge sampling change.


Figure 5: BLC1 NZO correctable term trends.

Using the lot groups as described above, we plot un-modeled BLC1 NZO signatures for X and Y in Figure 6.  As can be seen, the basic wafer level signature is fairly constant over the 5 months studied.  One exception is the small hot-spot that is apparent in group 2.  The exact root cause of this hot spot was not determined, but it has no impact since overlay modeling flyer (or outlier) removal eliminated those measurements from the process control scheme.  The only other apparent significant change is due to the increased sampling established for group 3 at the extreme edge of the wafer (indicated by green crescents) in Figure 6.  Figure 7 shows the same data plotted as vector maps by group.  For the wafer signatures, results are averaged by field.  For the field map, the results are averaged by location.  Again, the only major change over 5 months is due to increased extreme edge sampling, as discussed above.


Figure 6: BLC1 NZO wafer signatures by group for NZO X and NZO Y.


Figure 7: BLC1 NZO wafer signatures by group as vector plots.

For a more quantitative analysis of the BLC data, an NZO budget breakdown ANOVA analysis was performed, as described in the introduction.  For simplicity we focus on common (or static), lot-by-lot, and chuck-by-chuck variance.  The 23 lots of BLC NZO data, comprising 46 wafers (1 wafer per chuck of the dual scan lithography tool), 2265 points per wafer, and 25 points per field were analyzed in the 5D Analyzer budget breakdown tool as shown in Figure 8.  The results are shown in Figure 8 with variance versus component for raw data and various models for X and Y NZO.  The main conclusion is that the majority of the variance is static or “common” across the data set.  This is consistent with the analysis described above.  The second component, lot by lot, comprises significantly less variation.  As expected, chuck by chuck variance is insignificant for NZO.  The “error” is variance that is not captured in the other components.  These can include the other categories shown in Figure 2.

Figure 8 shows analysis of the raw data as compared with various models, including linear wafer and linear field (W1F1), high order wafer and high order field (W3F3), correction per exposure 2 (2 terms per field), correction per exposure 15 (15 terms per field), and correction per exposure 19 (19 terms per field).  It is apparent that the polynomial models and the low order CPE model do not properly capture the variance that is present in the data set, primarily due to the extreme wafer edge as will be shown below.  High order CPE can capture the edge variance, however, at the risk of noise sensitivity or over fitting.  It should be noted that the “error” or remaining variance of the raw data includes the residual model variance, whereas it does not for the modeled results.  The delta is an indication of the residual variance.

An equivalently large data set is available for BLC2.  The results and conclusions are similar, as expected, since they are very similar interleaved prints at the same layer.  For the sake of brevity, the BLC2 results are not shown here, with the exception of Figure 9, showing similar results to BLC1 in Figure 8: the common or static variance is dominant, indicating minimal NZO drift or excursion.


Figure 8: BLC1 NZO budget breakdown analysis by component.


Figure 9: BLC2 NZO budget breakdown analysis by component.

The edge versus the center of the wafer BLC1 NZO was analyzed.  The added extreme wafer edge sampling of BLC1 group 3 enables analysis of the extreme wafer edge behavior, as shown in Figures 10 and 11.  In Figure 10a, the polynomial W3F3 model does a reasonable job describing variability in the center of the wafer as seen by the relatively small residuals.  The extreme edge, however, is poorly characterized by W3F3 (red residual vectors > 2nm).  The CPE19 model manages to cover the extreme edge behavior, as seen in Figure 10b.  Figure 11 shows radial plots of the same data.  CPE19 modeling requires large sampling coverage, and can be susceptible to noise and over fitting.


Figure 10: BLC1 NZO group 3 edge analysis: (a) W3F3 residuals and (b) CPE19 residuals.


Figure 11: BLC1 NZO group 3 residuals as a function of wafer radius for W3F3 and CPE19: (a) residual X and (b) residual Y.

3. SCC Layer Analysis
Next we will describe the storage node contact layer SCC data analysis.  For this investigation, we sampled 18 lots, 1 wafer per chuck, for a total of 36 wafers at both DI and FI over a 4 month period.  The wafer sampling is somewhat reduced, as shown in Figure 12, at 378 points per wafer, and does not include extreme wafer edge points.  The field coverage is 25 points per field, however, in an asymmetric fashion (not all fields are sampled the same).  DI metrology is performed on Archer 500 on optical Archer AIM imaging targets.  FI metrology is performed by SEM on device structures as shown in Figure 12.  In this case only Y data is available for FI, due to the device structure.


Figure 12: SCC experimental setup (a) wafer sampling, (b) field sampling, (c) DI metrology with Archer AIM, and (d) FI metrology on the device structure.

The SCC NZO (delta between DI and FI overlay) trend charts are shown in Figure 13.  Keep in mind that the lots in this NZO study represent a small minority of total lots processed over the 4 month timespan.  Variability is present in the |mean|+3σ and NZO raw mean charts.  A change point is apparent; especially in the NZO raw mean (raw meaning not filtered, modeled, or residual), as indicated with the vertical line.  The origin of the change point is unknown, but is presumed to be a process tool preventative maintenance (PM) event.  Again we separate the results into groups.  Group 1 is before the change point, and groups 2 and 3 are an arbitrary split of all lots after the change point.  Figure 14 shows the wafer and field signatures before and after the apparent change point, as indicated by the horizontal line.  With the exception of the single change point, the wafer and field signatures remain fairly stable.


Figure 13: SCC trend chart showing NZO |mean|+3σ and NZO mean.


Figure 14: SCC NZO signatures by group: above wafer signatures for NZO Y, and below field signatures for NZO Y.

For a quantitative analysis of the SCC data, an NZO budget breakdown ANOVA analysis was performed, as described in the introduction.  Again, we focus on common (or static), lot-by-lot, and chuck-by-chuck variance.  The 18 lots of SCC NZO data, comprising 36 wafers (1 wafer per chuck of the dual scan lithography tool), 378 points per wafer, and 25 points per field were analyzed in the 5D Analyzer budget breakdown tool as shown in Figure 15.  The results show variance versus component for raw NZO data and various models for Y NZO.  In this case, due to the single change point, lot-by lot variance dominates over the static or common component.  As expected, chuck by chuck variance is insignificant for NZO.  The “error” is variance that is not captured in the other components.  These can include the other categories shown in Figure 2.

Figure 15 also shows analysis of the raw data as compared with various models, as discussed previously.  Since the extreme edge points are not included, the polynomial and CPE models are more comparable in their characterization of variance.  Again, the “error” or remaining variance of the raw data includes the residual model variance, whereas it does not for the modeled results.  The delta is an indication of the residual variance.


Figure 15: SCC NZO budget breakdown analysis by component.

The general picture is that while NZO includes some variability, it’s mostly static between change points.  The question is whether these change points can be identified quickly so that (1) NZO can be updated when necessary, but (2) not updated when it is not necessary.  One possibility is to use metrics from high frequency optical DI overlay metrology.  The DI optical metrology step includes information from both “current” and “previous” layer targets beyond overlay metrology.  The information can be encapsulated into metrics including Qmerit [2], as well as other quality metrics.  Figure 16, for example, shows an inflection in Qmerit for the SCC metrology, indicated for the vertical line, at the same time as the NZO inflection discussed above.  While more characterization work is required, this is suggestive that DI optical quality metrics which are already available at high frequency could be used to trigger NZO updates, thus reducing material at risk and reducing un-necessary expensive and time-consuming SEM FI metrology.


Figure 16: SCC DI Qmerit change point.

4. Conclusions
DI-FI bias, or NZO, has been incorporated into process control for many nodes.  What is new is the relative size of the NZO as compared to the overlay spec, and the need to find more comprehensive solutions to characterize and minimize the size and variability of NZO.  We implemented an analysis solution in 5D Analyzer which provides NZO corrected modeled terms to the APC host for the high frequency optical DI control loop based on low frequency updates to NZO based on DI minus FI SEM overlay results.

NZO signatures appear to be predominantly stable over ~5 month period, though variability exists on smaller scales.  Common or static signatures dominate NZO variance.  Lot to lot variance is present but not dominant, except for change points.  Chuck to chuck variance is observed to be insignificant.  Polynomial models characterize NZO inside ~140 nm wafer radius.  Extreme edge characterization requires CPE modeling, however, CPE is susceptible to over-fitting and noise.

Change points can be seen to alter the NZO signature, though for this study that occurred only infrequently.  What is needed is a methodology to identify change points in order to (1) minimize material at risk, and (2) minimize the cost of unnecessary NZO updates.  If inflections can be identified, low frequency FI sampling is adequate.  Already existing DI optical overlay metrics, such as Qmerit, show potential in identifying such change points by providing information on the previous and current level processing beyond overlay itself.

Future investigations include identifying NZO change points using DI metrics, investigations into NZO root causes, and the reduction of NZO and NZO variance with metrology optimization and metrology target optimization.

5. Acknowledgements
The authors would like to thank Onur Demirer, Fangren Ji, and Wayne (Wei) Zhou for help with the data analysis.

6. References
[1] Honggoo Lee, et. al., “Device overlay method for high volume manufacturing,” SPIE Volume 9778: Metrology, Inspection, and Process Control for Microlithography XXX, June 2016

[2] Dana Klein, et. al., “Quality metric for accurate overlay control in <20nm nodes,” Proc. SPIE 8681, Metrology, Inspection, and Process Control for Microlithography XXVII, 86811J (18 April 2013).

Initially published in SPIE Advanced Lithography Conference 2017: Honggoo Lee et al, “High-Volume Manufacturing Device Overlay Process Control,” Proc. SPIE 10145, Metrology, Inspection, and Process Control for Microlithography XXXI, 101450D (2017).



Leave a Reply