SMO Photomask Inspection in the Lithographic Plane Emily Gallagher1, Karen Badger1, Yutaka Kodera2, Jaione Tirapu Azpiroz3, Ioana Graur3 Scott D. Halle4, Kafai Lai3, Gregory R. McIntyre4, Mark J. Wihl5, Shaoyun Chen5, Ge Cong5, Bo Mu5, Zhian Guo5, Aditya Dayal5 1
IBM System & Technology Group, 1000 River Street, Essex Junction, VT, 05452 2 Toppan Photomasks, Inc., 1000 River Street, Essex Junction, VT, 05452 3 IBM Semiconductor Research & Development Center, Hopewell Junction, NY, 12533 4 IBM Systems & Technology Group, 257 Fuller Road, Albany, NY, 12203 5 KLA-Tencor Corporation, 160 Rio Robles, San Jose, CA, 95134 ABSTRACT Source Mask Optimization (SMO) describes the co-optimization of the illumination source and mask pattern in the frequency domain. While some restrictions for manufacturable sources and masks are included in the process, the resulting photomasks do not resemble the initial designs. Some common features of SMO masks are that the line edges are heavily fragmented, the minimum design features are small and there is no one-to-one correspondence between design and mask features. When it is not possible to link a single mask feature directly to its resist counterpart, traditional concepts of mask defects no longer apply and photomask inspection emerges as a significant challenge. Aerial Plane Inspection (API) is a lithographic inspection mode that moves the detection of defects to the lithographic plane. They can be deployed to study the lithographic impact of SMO mask defects. This paper briefly reviews SMO and the lithography inspection technologies and explores their applicability to 22nm designs by presenting SMO mask inspection results. These results are compared to simulated wafer print expectations. Keywords: Source Mask Optimization (SMO), Aerial Plane Inspection (API), lithography simulation
1. INTRODUCTION Traditional methods of extending lithographic resolution have become very difficult. Source Mask Optimization (SMO) has emerged as an attractive method of extending resolution on existing lithography scanners.[1,2] SMO adjusts the mask and source variables collectively to determine the optimum set of image-forming waves that can propagate within the finite NA of the exposure optics. Limits are imposed on the optimization solutions to ensure that the mask and source outputs are manufacturable. There are three broad areas of SMO development: algorithms, source and mask. Recent reports indicate significant progress on efficient algorithm development and complex source builds. This paper focuses on the mask component. With the luxury of mask shape constraints to ensure that the mask design can be manufactured, the focus is on the mask defects themselves. SMO masks have an extremely high density of edges and the correspondence between a mask edge and the target wafer edge is not clear. Traditional concepts of a mask defect must be reconsidered when the one-to-one correspondence between design and mask features is lost. This is most easily done with programmed defects added to the design data. Simulations are used to establish the defect printability on wafer. A mask using the same design was built and inspected on a KLA-Tencor 597XR optical inspection tool. Conventional Reticle Plane Inspection (RPI) and Aerial Plane Inspection (API) methods were applied to the SMO mask. High resolution inspection is one option, but the API simulation offers the advantage of applying wafer-level detection limits. This wafer view will provide a more manufacturable solution since only defects that print are identified. To avoid questions about the validity of resist model calibration, both the simulations and the inspection results were analyzed in the aerial plane.
Photomask Technology 2009, edited by Larry S. Zurbrick, M. Warren Montgomery, Proc. of SPIE Vol. 7488, 748807 路 漏 2009 SPIE 路 CCC code: 0277-786X/09/$18 路 doi: 10.1117/12.830668
Proc. of SPIE Vol. 7488 748807-1 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
The goal of this work is two-fold. The first goal is to demonstrate that mask inspection is possible on the small structures found on SMO masks. The second goal is to understand how a mask’s defect and its location influence the SMO wafer contour. The SMO pattern deployed for inspection work shows that location is a good predictor of influence. This result was not for SMO because of the global influence of each mask feature.
2. SOURCE MASK OPTIMIZATION 2.1 Pattern Selection An immature 22nm contact level SMO mask solution was selected as the base design for this work. It had relatively low contrast and consequently is likely to show changes to the wafer contour at nominal dose and focus. The negative side of this choice is that print-through and line edge roughness are likely to be high on actual wafer prints. These wafer level effects prevented us from performing careful measurements on wafer prints. The pattern in Fig. 1 shows the high number of edges characteristic of SMO mask patterns. The blue features are clear regions. At 4X mask dimensions, the jogs can be as small as 4nm and the minimum mask feature size is ~50nm. These can be constrained further during using manufacturability constraints if desired. The mask process used for the test mask used a 50keV e-beam mask writer and OMOG binary substrate. The fidelity is quite good as can be seen by the pattern match of the CDSEM superimposed on the design. The smoothing of the mask images is relatively low with this process, but the effect is visible on small jogs.
Fig. 1. The SMO SRAM pattern selected for simulation and inspection. Blue shapes are clear on the mask. An SEM of the final mask is superimposed on the layout.
2.2 Programmed Defects Defects were inserted systematically into the design to study printability, inspectability and sensitivity. While many defect types were included, this paper concentrates the distributed defect categories: shape bias and edge movements. Extensions and intrusions were also studied, but because the area involved is smaller, the impact to wafer is less. The programmed defects were inserted into the middle of an SRAM array as shown in Fig. 2. Schematics of the nine defect types studied are also shown. Each defect type was sized from -10 nm to +10nm in 1nm increments in the data (1X).
Proc. of SPIE Vol. 7488 748807-2 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
Programmed Defect Types
#1 Bias
#2 Bias
#3 Bias
#4 Bias
#5 Move
#6 Move
#7 Move
#8 Move
#8 Move
Fig. 2. Programmed defects were added near the center of the base pattern as is indicated on the pattern (left). This example is an over-sized bias of the entire feature, called defect type #1. The nine different programmed defect types studied here are shown (right).
The defect sizing on the test mask must be well-matched to the design targets for comparisons between simulations and mask inspection to have integrity. CDSEM measurements of defects were taken using standard 1D measurement gates. The measurement results for defect type #2 shown for two masks using slightly different mask processes are shown in Fig. 3. These 1D SEM results are very close to target with errors that are typically < 1nm at 4X. Design data is appropriate input with modification for simulations; however, this is clearly an error source. Defect #2: Measured vs. Design Size 30 25
Measured (nm@4X)
20 15 10 5 -30
-25
-20
-15
-10
-5
0 -5 0
SMO-I 5
10
15
20
25
30
SMO-II
-10 -15 -20 -25 -30 Design (nm@4X)
Fig. 3. Programmed defects were measured relative to target on a CDSEM on two different masks. The measured result is plotted vs. design target for defect type #2.
3. SIMULATION 3.1 TMA and EMF Robust printability of designs over the entire process window is a critical and many simulation tools have evolved to help.[3,4] Full designs are simulated in the wafer plane to search for areas of weak printability so that they can be fixed. We are using it here to study the areas with programmed defects and measure the severity. Analyzing the impact of variations in mask shapes is a measure of co-optimization of solutions, SMO or conventional RET. Two general methods for simulation were used. The simplest, fastest method uses a Thin Mask Approximation (TMA). The mask absorber is assumed to have zero thickness and is characterized by transmission and phase shift only. Transmission and phase are then independent of the angle of incidence and polarization is not affected by the mask.
Proc. of SPIE Vol. 7488 748807-3 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
More rigorous methods approximate Electromagnetic Field effects (EMF) and typically increase simulation runtime relative to TMA methods. We used Mentor Graphicsâ&#x20AC;&#x2122; Domain Decomposition Method (DDM) to provide a more realistic representation of the impact of absorber topography. The absorber films are characterized by n, k and thickness, not phase and transmission. Transmitted field amplitude and phase depend on the feature dimensions, illumination angle of incidence and polarization. EMF effects induce differences between TMA and EMF on binary blanks: spaces in the mask appear smaller and feature-dependent shifts in best focus occur. The DDM technique relies on storing the electromagnetic fields diffracted by an isolated edge of the mask topography in memory [5]. These fields are simulated on a plane parallel to the mask surface and in close proximity to its exit surface through a rigorous numerical solution of Maxwell equations. This diffracted edge field is then applied to every edge of the mask design during simulation of the aerial image. This approximates the complex interaction between the illumination and the mask topography. An even more rigorous computation of the mask effects used the finitedifferences time-domain code implemented on Blue Gene clusters [6]. The resulting aerial image contours at the wafer plane are shown in Fig. 4 and are very similar to those generated with DDM, providing confidence on the adequate accuracy of the DDM results used in this paper. For all cases, the wafer contours are shown at nominal conditions using a constant threshold resist model anchored to print contacts 50nm (1X).
FDTD vs TMA
DDM vs TMA
(b)
(a)
Fig. 4. Wafer contours obtained with a constant threshold resist model anchored to print contacts with 50nm wide at nominal focus conditions: (a) TMA vs. rigorous simulation of the fields diffracted by the photomask with finitedifferences time-domain Maxwell solver and (b) TMA vs. conventional full-DDM model contours
3.2 Printability The simulation tools offer a wealth of analysis options. It is important to choose an analysis method that mimics the inspection tool simulations. The inspection simulator uses nominal dose and focus, so we restricted the TMA and EMF simulations to nominal as well. Fig. 5 shows the base SRAM pattern with the simulated contact contours in yellow. There is a characteristic lack of similarity between the wafer images and the mask design shapes in this SMO layout. Printability of a defect is determined by simulating the base pattern in the aerial plane at nominal dose and focus for the reference. Then the same simulation is repeated with the defect inserted. Both are anchored to a 50nm contact width at 1X. If there is a difference between the two simulations, there is a printing impact. This technique was repeated for both TMA simulations and EMF simulations.
Fig. 5. Printability of a defect is determined by simulating the base pattern in the aerial plane at nominal dose and focus for the reference. The resulting contours are shown superimposed on the base design.
The simulation printability of defects is displayed in a sensitivity chart format borrowed from inspection methods. The defect sizes are in rows with no defect indicated by the center line. Increasingly larger defects are above the line and increasingly smaller defects are below. If the simulation with defect differed from the simulation without defect at
Proc. of SPIE Vol. 7488 748807-4 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
nominal does and focus the cell is shaded. The defect types are ordered by columns. Fig. 6 shows results for both TMA and EMF.
Defect 2
Defect 3
Defect 4
Defect 5
Defect 6
Defect 7
Defect 8
Defect 9
EMF
TMA
EMF
TMA
EMF
TMA
EMF
TMA
EMF
TMA
EMF
TMA
EMF
TMA
EMF
TMA
EMF
Data Source
TMA
Oversize Undersize
Defect size (nm 1X)
Defect 1 +10 +9 +8 +7 +6 +5 +4 +3 +2 +1 -1 -2 -3 -4 -5 -6 -7 -8 -9 -10
Fig. 6. Printability of defects displayed in a sensitivity chart format borrowed from inspection methods. The defect sizes are in rows with no defect indicated by the center line, increasingly larger defects above the line and increasingly smaller defects below. If the simulation with defect differed from the simulation without defect at nominal does and focus the cell is shaded. Results are shown by defect type for TMA (left, solid) and EMF (right, hatched).
In general, TMA and EMF defect printability results are very similar for this binary mask. This is expected since relative differences between contours with and without defects are likely to be similar whether the contours are shifted by EMF, or not. Even when small discrepancies between this relative difference for TMA and EMF exist, the measurement is still sensitive with similar accuracy to the presence of a defect in the mask. It is the relative variation that is important. The anticipated result of using full EMF simulations is that the mask topography is treated so that spaces act smaller than they would with TMA simulations. Only four defect types show any difference in printability, and those differences are subtle. From an inspection perspective, TMA simulations appear to be a sufficient. 3.3 Defect Classification The simulation results revealed a surprising pattern: the printability of a defective feature depends heavily on where it is located relative to the wafer contour. Features fall into three general categories: 1. Primary features are mask features that are located directly under a contour. Defects on these features are likely to print even at very small sizes. Both TMA and EMF simulations predict the same printability. 2. Secondary features are mask features that are located near a contour. Defects on these features are likely to print when they are over-sized, but have a low sensitivity to defects when they are under-sized. Their response is asymmetric. TMA and EMF predict slightly different printability. 3. Tertiary features are mask features that are located farthest from the contour. Defects on these features are unlikely to print even at large sizes. Both TMA and EMF simulations predict the same printability. This suggests that preconceived notions of assists and primary features do not have to be abandoned for SMO, they simply need to be modified for the more complex patterns.
4. INSPECTION 4.1 Inspection Overview API is an inspection mode offered by KLA-Tencor. API moves the inspection to the wafer aerial plane with the several associated advantages over the reticle. Restrictions placed on designs for inspectability can be relaxed. Nuisance defects are filtered so that only printable ones are detected. [8,9]. The AIMS verification analysis is significantly reduced.
Proc. of SPIE Vol. 7488 748807-5 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
The inspection steps are shown schematically in Fig. 7. There are two paths for test and reference. The first step of the test path is to collect the high resolution images of the mask and reconstruct a physical model of the mask pattern, including any pattern defects. This is called mask pattern recovery. There is no need to be actinic at this point since the mask pattern recovery process depends on high resolution images, not the wavelength of capture. The first step of the reference path starts from the design and simulates what the mask would look like. From that point the test and reference paths use the same lithographic simulations to generate an aerial image, but with the different mask inputs. An optional final step applies a resist threshold to the aerial image.
Fig. 7. A schematic of the inspection method shows two data paths. The test is simulated from transmitted and reflected inspection images in the reticle plane (RPI). The reference is simulated directly from the design. Comparisons can be made at either the aerial plane or at the wafer plane.
Detection occurs when the test and reference aerial plane images are subtracted. If the 2x2 difference intensity falls between the upper and lower thresholds, the defect is registered. The detection methodology is shown in Fig. 8.
Fig. 8. Detection method in the aerial plane is illustrated by plotting the image intensity vs. position of a test and reference image. A filter is applied by imposing an upper and lower limit to the intensity. If there is a difference between the test and reference image and the difference is within the limits chosen, then the defect is considered real.
4.2 Inspection Results The inspection data is presented in two sensitivity charts. Fig. 9 compares conventional inspection (RPI) results to the API results. For inspection sensitivity to be indicated on the chart the inspections were run 10 times and the defect had to be captured in all runs. RPI inspection ran well with reasonable sensitivity and no false detections. API also ran well, but with better sensitivity. How do these real mask inspections compare to the defect printability as predicted by the TMA simulation?
Proc. of SPIE Vol. 7488 748807-6 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
API
Defect 9
RPI
API
Defect 8
RPI
API
Defect 7
RPI
API
Defect 6
RPI
API
Defect 5
RPI
API
Defect 4
RPI
API
Defect 3
RPI
API
Defect 2
RPI
API
Data Source
Defect 1
RPI
Oversize
Defect size (nm 1X)
Undersize
+10 +9 +8 +7 +6 +5 +4 +3 +2 +1 -1 -2 -3 -4 -5 -6 -7 -8 -9 -10
Fig. 9. Printability of defects displayed in a sensitivity chart format. The defect sizes are in rows with no defect indicated by the center line, increasingly larger defects above the line and increasingly smaller defects below. The RPI results are shown by defect type for RPI (left, solid) and API (right, hatched).
API inspection sensitivity is compared to TMA printability in Fig. 10. It is important to remember that the API defect inspection was not run at the maximum sensitivity, but detuned to levels appropriate for real masks. Generally the TMA simulation printability occurred before detections were registered on the real masks. This is expected. The API detection is ~ 6-8% CD error as calculated using a traditional 1D CD measurement along the long axis of the simulated contact. While the CD analysis method is familiar, it is inadequate since many of these defects result in CD errors on multiple contours along multiple axes. A more comprehensive analysis method is needed.
Defect 2
Defect 3
Defect 4
Defect 5
Defect 6
Defect 7
Defect 8
Defect 9
API
TMA
API
TMA
API
TMA
API
TMA
API
TMA
API
TMA
API
TMA
API
TMA
API
Data Source
TMA
Oversize Undersize
Defect size (nm 1X)
Defect 1 +10 +9 +8 +7 +6 +5 +4 +3 +2 +1 -1 -2 -3 -4 -5 -6 -7 -8 -9 -10
Fig. 10. Printability of defects displayed in a sensitivity chart format. The defect sizes are in rows with no defect indicated by the center line, increasingly larger defects above the line and increasingly smaller defects below. The results are shown by defect type for TMA (left, solid) and API (right, hatched).
A new overlay method is introduced to compare TMA and API since a single CD error is inadequate. The TMA binary image and API grayscale images had to be converted. The API images were converted to binary and up-sampled from 9nm/pixel to 2.25nm/pixel. The TMA images were already binary, but had to be down-sampled from 1.4nm/pixel to 2.25nm/pixel. The two pixel-matched images were overlaid. Common areas were determined for each shape, normalized to an average reference and summed over all shapes to calculate a pattern matching score as shown in Fig. 11. A defectfree pattern registers a score of 97.7%. Examples with defects exhibit pattern matching scores that are also very high. This is currently a manual process, but points to a method for image comparison that is more comprehensive than CD error.
Proc. of SPIE Vol. 7488 748807-7 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms
all shapes
Fit (%) =
Σ (A i =1
com i
/ Aref i) X 100
Fig. 11. TMA and API contours are overlaid for matching analysis. The under-sized contact is seen in both TMA and API simulations. A calculation of the matching determined that this defect had a pattern matching factor of 95.5%.
5. CONCLUSIONS This paper explores defects in an SMO mask design, how they behave and if they can be detected at mask inspection. The relationship between defects and their location in SMO patterns was determined with EMF and TMA simulations. The printability of sizing defects fell into three categories: primary ones that occur under the wafer contour, secondary ones that occur near the wafer contour and tertiary ones that are farthest from the contour. The impact of the defect is linked with its location on the mask relative to the wafer contour. Sensitivity and inspectability was good on all SMO defect types studied using both Reticle Plane and Aerial Plane inspection modes. Despite the challenges introduced with SMO patterns, mask inspection was demonstrated.
ACKNOWLEDGEMENTS The authors would like to thank the KLA-Tencor RAPID WPI engineering and application teams for their development of the API technology and the IBM manufacturing and engineering teams for mask processing.
REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9]
K. Lai, et al., “Experimental Result and Simulation Analysis for the Use of Customized Illumination from Source Mask Optimization for 22nm Logic Lithography Process,” Proc. SPIE 7274, (2009). A.E. Rosenbluth, et al., “Optimum Mask and Source Patterns to Print a Given Shape,” JM3, 13 (2002). S. Mansfield, et al., “Lithography simulation in DfM: achievable accuracy versus requirements” Proc. SPIE 6521, (2007). J. A Torres and C.N. Berflund, “Integrated circuit DFM framework for deep sub-wavelength processes” Proc. SPIE 5756, 39 (2005). K. Adam and A. Neureuther, “Methodology for accurate and rapid simulation of large arbitrary 2D layouts of advanced photomasks,” Proc. SPIE 4562, 1051–1067, (2002). J. Tirapu-Azpiroz, G.W. Burr, A.E. Rosenbluth, and M.S. Hibbs, “Massively-Parallel FDTD Simulations to Address Mask Electromagnetic Effects in Hyper-NA Immersion Lithography,” Proc. SPIE 6924, (2008). K. Adam, "Modeling of Electromagnetic Effects from Mask Topography at Full-Chip Scale," Proc. SPIE 5754, (2005). C. Hess, M. Wihl, R.-f. Shi, Y. Xiong, S. Pang, “A Novel Approach: High Resolution Inspection with Wafer Plane Defect Detection,” Proc. SPIE 7028, (2008). E. Gallagher, et al., “Wafer Plane Inspection Evaluated for Photomask Production”, Proc. SPIE 7128, (2008).
Proc. of SPIE Vol. 7488 748807-8 Downloaded from SPIE Digital Library on 18 Nov 2009 to 192.146.1.254. Terms of Use: http://spiedl.org/terms