Improvement of silicon waveguide transmission by advanced e-beam lithography data fracturing strategies

Transcription

1 Improvement of silicon waveguide transmission by advanced e-beam lithography data fracturing strategies Running title: Improve silicon waveguide transmission by fracturing strategies Running Authors: N. S. Patrick, R. J. Bojko, S. J. H. Stammberger, E. Luan, L. Chrostowski N. Shane Patrick a) University of Washington, Washington Nanofabrication Facility, Electrical Engineering, Box , Seattle, Washington GenISys Inc., San Francisco, CA GenISys GmbH, Munich, Germany Richard J. Bojko Stefan J.H. Stammberger Enxiao Luan, Lukas Chrostowski University of British Columbia, Vancouver, Canada a) Electronic mail: In the maturing field of silicon photonics, advances continue in both design and process improvements. Waveguide propagation loss is strongly affected by sidewall roughness, so for fabrication using e-beam lithography, loss is influenced by e-beam writing parameters. Here, we look specifically at fracturing strategies in data preparation for e-beam lithography, and find significant reduction in waveguide loss by utilizing advanced fracturing options. 1

2 For our evaluation, we fabricate optical waveguides using a well-characterized, highlystable baseline fabrication process with HSQ resist exposed by a 100 kv electron beam, a high-contrast TMAH develop, and a Cl2 ICP etch. Using surface grating couplers for input and output, automated optical measurements are made by scanning input light in the region of the design wavelength of 1550 nm and measuring optical output power. We use a design cell containing grating couplers and both straight and curved waveguides with a range of lengths. We find significant reduction in grating coupler insertion loss and waveguide loss along with increased uniformity by leveraging a new fracturing strategy implemented in the BEAMER pattern data processing software from GenISys, GmbH. Single Line Edge Smoothing (SLS) is an exposure strategy in which all feature edges are traced using a single-line shape (sometimes referred to as a single-pass line) while the bulk of the shape is then exposed with trapezoidal beam filling. The insertion loss for grating couplers written using Single Line Edge Smoothing shows a significant loss reduction of 1.2 db as well as greatly improved uniformity. Both straight and curved waveguide losses were also reduced by use of Single Line Edge Smoothing, by 0.7 and 1.1 db/cm respectively. Here, we will discuss the likely mechanisms of this improvement as well as present additional device data using these new fracturing methods which represent a significant, incremental improvement in performance of optical waveguides written by e-beam lithography. 2

3 I. INTRODUCTION Silicon photonics is a category of devices in which light traveling through silicon is utilized to convey and process data. It can be thought of as the direct analogue for working with light as standard semiconductor circuits are for working with electrons. Circuit traces are replaced by waveguides, wire pads are replaced by couplers, and analogues even exist for creating basic logic devices. As silicon is widely established in the industry, using it for research into new methodologies often allows the use of existing technology and infrastructure, thus reducing costs and making this an attractive area for faster computing research. Working with light, as compared to electrons, has a unique set of challenges even at the most basic of levels. Relevant to this work, electron confinement and transmission across metal traces, and even arcing in various materials, is well-characterized, but light travelling as a wave is more inclined to travel through material boundaries rather than going along its intended path within a material. Insertion and propagation loss of light is described as the relative amplitude of the light at the point of interest versus the amplitude at the starting point and is often described in decibels (db) or decibels per centimeter (db/cm). These numbers provide a quantitative description of how well the light is confined to the silicon structures. Any light that is not confined diffuses into the material surrounding the silicon features of interest. Thus, all structures are designed so that when light interacts at a material boundary it is reflected appropriately to achieve total internal reflection without scattering. As with any interaction between light and a mirror, the smoother the surface the better the reflection. 3

4 Silicon propagation loss, then, is strongly affected by sidewall roughness of all involved silicon fabricated features. In fabrication using e-beam lithography, wherein an electron beam is swept across the material to draw the desired structures directly, loss can be reduced by optimizing writing parameters, including fracturing strategy. E-beam lithography using a vector-scan, Gaussian-beam system requires the designed patterns to be fractured into fundamental shapes, typically rectangles and trapezoids. As e-beam writing in such a system is done on a grid, both straight and curved structures can suffer from this fracturing necessity resulting in feature edges with non-ideal shot placement. Here, we investigate a new fracturing method that allows tracing these features with an additional shape to allow more precise edge placement of shots. Applying this new strategy available in the BEAMER pattern data processing software to the established WNF lithography and etching processes for silicon photonics, 1 we can achieve the expected process fidelity along with significant reduction in grating coupler insertion loss and waveguide loss along with improved uniformity. II. EXPERIMENTAL SETUP AND METHODOLOGY A. Fabrication Parameters Waveguides were fabricated using the baseline WNF process. 1 The substrate was silicon-on-insulator (SOI) with an active silicon (Si) layer thickness of 220 nm and a bottom oxide (BOx) thickness of 3000 nm, diced into 25 mm squares. Patterning was performed using hydrogen silsesquioxane (HSQ) resist exposed in a JEOL JBX-6300FS electron beam lithography system using 100kV beam energy, 8 na beam current, base dose of 1800 μc/cm 2, dose-modulation proximity effect correction, and a field-shift of 2, 4

5 in which the pattern is written twice with a shift of one-half of the field size of 500 μm between the two written passes. The fundamental placement grid was 1 nm; this is the minimum increment to which a shape can be positioned. However, to achieve reasonable throughput for large patterns, a pixel spacing of 6 nm was used; this is also known as the shot-pitch, or beam step size, and is the spacing between pixels filling each shape. Resist development was in 25% tetramethyl-ammonium hydroxide (TMAH) for 4 minutes, followed by rinsing first in deionized (DI) water, then isopropyl alcohol (IPA), and finally drying with nitrogen (N2). The pattern was then transferred into the silicon device layer using a chlorine (Cl2) Inductively-Coupled Plasma (ICP) etch. The remaining resist was then stripped in buffered hydrofluoric acid (HF), and a passivation layer of 2000 nm of SiO2 was deposited using PECVD. B. Test Patterns 1. Grating Couplers To measure grating coupler insertion loss, two subwavelength surface grating couplers 2 (Fig. 1) are connected via the shortest length of waveguide possible for the testing setup (Fig. 2). The transmission between the two is measured and the propagation loss in the short segment of wavelength is considered negligible. This coupler is designed to operate at 1550 nm, TE00 mode, with a 25 degree entrance angle. The physical design of the subwavelength grating coupler used consists of a primary grating of 556 nm period and 260 nm width, with subwavelength grating features of 80 nm. Complete design details, modeling, design sensitivity results, and performance results are in the reference provided 2. 5

6 FIG. 1. Subwavelength surface grating coupler as designed in CAD software. FIG. 2. CAD layout for grating coupler insertion loss testing featuring 2 subwavelength grating couplers spaced a fiber-width apart to accommodate testing setup connected by the shortest possible length of waveguide. Curvature is designed to not exceed critical angle above which light is more likely to scatter into the surrounding material. 2. Waveguides To measure waveguide propagation loss, straight waveguides in 5 lengths from 1882 μm to μm, and curved waveguides in 5 lengths ranging from 2381 μm to 8505 μm were fabricated. (Fig. 3) All waveguides are nominally 500 nm wide by 220 nm high, and have a slightly trapezoidal cross-section due to an 82 degree sidewall slope from the ICP etching process previously described 1. A minimum bend radius of 5 μm is 6

7 used throughout the design to ensure no excess bend losses occur. Light is coupled into and out of these structures through the subwavelength grating couplers detailed above. FIG. 3. CAD layout for waveguides showing curved and straight designs of varied lengths. Curved segments at ends of straight guides are for design compaction only and are assumed to have a negligible impact. 3. Ring Resonators As an additional characterization of propagation loss, ring resonators (Fig. 4) were fabricated for examination and comparison of their quality (Q) factors. Light is inserted into a waveguide segment that runs in close proximity to the ring structure. As the light travels through the waveguide, a portion of it couples through the gap into the resonant structure. This light travels around the ring, coupling into a waveguide on the opposite side of the structure and continuing around to couple back into itself to create constructive and destructive interference at particular frequencies depending on design parameters. In this case, the interference pattern is of less concern than how long the resonance continues without input, which we measure by extrapolating the Q-factor. For a resonant system, a higher Q-factor indicates that the oscillations persist longer which 7

8 means the intrinsic loss of energy is lower. Since these rings are essentially curved waveguides, this allows another avenue to confirm smoothness improvements. FIG. 4. CAD layout for ring resonators showing the insertion couplers (connected) and the resonance measurement coupler (single coupler only). C. Data Preparation All data preparation was performed using BEAMER, a software suite created and maintained by GenISys GmbH. The same structures were fabricated twice, side-by-side once using the baseline WNF process and once using the new advanced fracturing strategies (SLS) as a modification to the baseline WNF process. 1. Baseline WNF Method There are many pieces to building a successful pattern process flow, however only the fracturing and tool-specific conversion is being investigated here. For more information on the overall process, see the referenced publication. 1 The standard fracturing method uses a user-specified beam placement (shot) spacing referred to as the beam step size or shot pitch. The choice of this spacing is dependent on numerous factors including beam current, feature size, tool clock speed, and other often tool-specific considerations. For this work, a beam step size of 6nm was used for higher throughput, per our established standard process, but the fundamental shape placement is at the 1 nm machine grid. The conventional shape fracturing method was employed wherein the software attempts to break the shape into as few rectangles 8

9 and trapezoids as possible while being as accurate as possible. Shots are then placed within these shapes according to a grid. Exact placement is tool-specific, but here the shots are placed on a 1nm grid and the tool offsets the grid lines for better edge accuracy. (Fig 5) FIG. 5. Illustrative waveguide (not to design size) at an arbitrary angle, standard fracturing (shapes not shown), with simulated shot placements shown. Shots are not shown at full size for clarity of illustration. Note that shot placements on the edges approximate the actual shape edge, but with significant alternations that contribute to edge roughness. 2. Experimental Single Line Edge Smoothing (SLS) Method The experimental fracturing method follows the same general flow as the baseline method, but modified to give extra attention to edge definition. First, prior to fracturing, all features are traced with a new polygon that is exactly one beam step size wide (6nm). This new shape is set 2/3 of a beam step size inwards from the line edge and overlaps the original feature by 1/3 of a shot pitch. To accomplish this, the original shape is biased down by 2/3 of a shot pitch, thus retaining the original design dimensions. This portion of the methodology is tool-agnostic and can be used across any EBL instrument. These steps 9

10 are automated within the BEAMER data processing flow, and the initial implementation of having an overlap between the single line and the bulk-fill shapes of 1/3 of a shot pitch was selected to provide reasonable dose continuity between the shapes, without excessive gaps or overlaps. The slight overlap of the Gaussian beams also provides an increase in the edge contrast of the exposure intensity. This overlap could be varied in future studies, if desired. Standard dose-modulation proximity effect correction was then applied. Since the single line shapes are contiguous with the filling trapezoids, the doses assigned to the single line shapes were identical to the immediately adjacent trapezoid shapes, although another variation for possible future exploration could be to artificially further modulate the single line doses, which might have additional effects on the line-edge roughness, depending on the resist response to dose. Making the best of the fractured data, as with any method, requires some toolspecific knowledge and formatting. In this case, the data is formatted by BEAMER to take advantage of an additional primitive shape of the JEOL JBX-6300FS system known as the single line shape, sometimes called a single-pass line. Unlike trapezoids, which are specified as a set of vertices and then filled with shots spaced by the exposure grid (6 nm in this work), single line shapes are specified as a starting point, a length, and an angle with shots placed in a straight line adhering to the same beam step size, but able to position each pixel the nearest point on the machine grid of 1 nm. An additional small dose modulation is automatically applied by the e-beam tool itself to account for increased pixel spacing due to the rotation angle of the single line. During the BEAMER export to the JEOL-specific pattern format, all of the single beam step size wide 10

11 trapezoids created earlier are converted into these single-line primitives for writing, providing a vector scan parallel to the shape edges, acting like a pencil to trace the edges of our shapes cleanly and smoothly. (Fig. 6) In this work, the patterns were processed into the data format specific to the JEOL machine. Some other direct-write e-beams also have similar single-line primitives, although the exact details and capabilities vary with each tool vendor. FIG. 6. Illustrative waveguide (not to design size) at an arbitrary angle, SLS fracturing (shapes not shown), with simulated shot placements shown. Shots are not shown at full size for clarity of illustration. Note that shot placements on the edges approximate the actual shape edge more closely than the standard fracturing shot placement by placing each shot on the 1 nm machine grid instead of the 6 nm shot pitch. D. Automated Measurements Measurements were conducted using an automated set of hardware and software 3, allowing for the testing of a far higher number of devices than would be possible with manual measurements. The system used consists of a four-fiber linear array, motorized X-Y stage, manual z-focus, and computer interface. The device uses information in the design CAD file to locate each set of input/output grating couplers, confirming the 11

12 location by doing a small spiral outwards while monitoring for maximum intensity returned which is assumed to be the appropriate measurement site. Light is then swept in a small range around the designed wavelength (here, 1550nm). Optical output is recorded by discrete sampling and stored in a text file for processing. (Fig. 7) For each device, the peak transmission value was compared to calibration structures to determine attenuation in db. Using this setup, over 1300 individual devices were measured for this experiment. E. Data Analysis Grating coupler insertion loss in db is determined from the transmission spectrum of the simple grating coupler loop described above by simply halving the total loss at the peak transmission wavelength, since the input and output will be reciprocal. A typical grating coupler transmission measurement result is shown in Figure 7a. For waveguides, the transmission loss in db/cm is determined by the slope of a linear regression of the measured peak waveguide transmission versus the waveguide length, with a typical family of transmission curves shown in Figure 7b, where the five transmission curves represent the five different lengths measured. Regression plots for straight and curved waveguides are shown in Figure 9. A representative transmission spectrum for a ring resonator is shown in Figure 7c. The Q-factor for a ring resonator is determined by fitting the central resonance peak with a Lorentzian function and then dividing the peak wavelength by the peak width. Interpretation of this value is complicated in that it has 12

13 contributions from both coupling and intrinsic waveguide loss, which are not completely separable, thus only total Q-factor change is reported. FIG. 7. (Color online) Example of data collected from automated measurements and displayed as a line graph of transmission (db) versus wavelength (nm). Examples are shown are characteristic for (a) grating coupler, (b) waveguide, and (c) ring resonator test structures. The multiple lines in (b) represent the different lengths of waveguides being measured. III. RESULTS A. Grating Coupler Insertion Loss 13

14 Grating couplers fabricated by the baseline method show a mean insertion loss of -8.99dB and a standard deviation of 0.43dB. Grating couplers fabricated by the experimental SLS method showed a mean insertion loss of -7.77dB with a standard deviation of 0.12dB. The SLS method, then, provides a mean loss reduction of approximately 1.2dB and uniformity improvement of approximately 3 times over the baseline. (Fig. 8). FIG. 8. (Color Online) Box and whisker plot showing grating coupler insertion loss values for both the baseline and SLS fracturing methods. B. Waveguide Transmission Loss Linear waveguides fabricated by the baseline method show a mean transmission loss of -1.9 db/cm with a standard error of 0.4 db/cm across both fabrication runs, while curved waveguides show a mean loss of -6.0 db/cm with a standard error of 0.6 db/cm. Linear waveguides fabricated by the experimental SLS method showed a mean transmission loss of -1.4dB/cm across both fabrication runs with a standard error of

15 db/cm, while curved waveguides show a mean loss of -4.5dB/cm with a standard error of 0.7 db/cm. The SLS method, then, displays a mean loss reduction of approximately 0.5dB/cm over the baseline for straight waveguides, and a reduction of 1.5dB/cm for curved waveguides. (Fig. 9). FIG. 9. (Color Online) Line graphs showing transmission (db) versus design length (µm) for both linear and curved waveguides for both baseline and SLS fracturing methods across both production runs studied. C. Ring Resonator Q-Factor Ring resonators fabricated by the baseline method show a mean total Q-factor of with a standard deviation of 4909, while those fabricated by the experimental SLS method showed a mean total Q-factor of with a standard deviation of The SLS method displays an improvement of approximately 27% over the baseline. (Fig. 10). 15

16 FIG. 10. (Color online) Box and whisker plot showing ring resonator total Q-factor values for both the baseline and SLS fracturing methods. IV. DISCUSSION While the data indicates that the SLS method of fracturing data does indeed provide improvements in silicon photonics devices, there are multiple factors that may contribute to the observed improvements. A. Line Edge Roughness The foundational knowledge going into this experiment is that line edge roughness (LER), or sidewall roughness, is known to have a dramatic impact on silicon photonics losses due to the simple fact that rough surfaces make poor mirrors. While every step in a fabrication process like this one has its own contributions to this line edge roughness, we have shown that for fabrication using electron beam lithography, fracturing strategies do impact how smoothly the shapes are drawn in the first place. Unfortunately, reliable direct measurement of line edge roughness was not possible with 16

17 the equipment available. However, as a simple example, rudimentary analysis of shot placements for various features, such as those shown in figures 5 and 6, predicts a theoretical reduction in the portion of line edge roughness attributable to shot placement from 4 nm to 1 nm. Of course, this simplistic analysis is not representative of an entire curved structure, but is provided only as an illustrative example of the LER reduction from the SLS method by the improved shot placement resolution of the single line edge features. This improved pixel placement of the SLS fracturing cannot in itself explain the observed improvement in linear waveguides, however, since the pixel placement on edges parallel to either of the scan axes is unchanged. B. SLS Overlap As described above, there is an overlap region of 1/3 of a beam step size between the single line features and the bulk of the designed structure after processing. Overlapping features represent zones that will receive higher doses due to being exposed by the electron beam more than once. The effects of this overlap is not established, but may explain some measure of the improvement observed in linear waveguide transmission loss. C. Critical Dimension Change Differences in feature sizes affect the performance of photonics structures. Since the pattern data is being directly biased as part of the SLS method, and the two shapes are written separately, there is some question over whether final dimensions are identical between the two fracturing methods. Direct SEM measurements of the samples studied here were inconsistent, but indicated a feature size growth of about 5 nm when using SLS 17

18 fracturing. Using the device measurements to calculate the group index values 4 and then estimating the linewidth change required to cause that change in group index predicts a growth of 7 nm. This feature growth for the SLS method likely contributes somewhat to the improved transmission loss, but generally width variations mostly affect the center frequency of the light that is confined to the structures. Collectively, these effects likely explain the observed performance improvements. The waveguide transmission improvement is well-explained by the expected reduction in line edge roughness from using single line edge smoothing. The uniformity improvement in the grating couplers is, however, particularly striking. This improvement can likely be attributed to the better shape edge fidelity and reduced line edge roughness of the subwavelength feature of the grating coupler, which is an 80 nm linewidth by far the smallest feature size in this study. Prior process characterization has shown the minimum viable feature size for this process is about 70 nm, so the 80 nm design point is not far from the process limit. It is not unexpected, then, that features this close to an edge condition should be more sensitive to even small process changes. V. SUMMARY AND CONCLUSION Despite some unknowns and some unquantifiable aspects of processing, the large sample size and consistency of general trend between production runs is clear. Utilizing electron beam lithography, silicon photonics losses are decreased by choice of fracturing strategies. New strategies utilizing SLS successfully reduce these losses by reducing sidewall roughness. 18

19 Opportunities for further investigation of this method include deeper investigation into the effect of the local dose change near the feature edge due to the small overlap introduced between the single line shapes and the filling trapezoids, the application to other devices such as Mach-Zehnder interferometers and photonic crystals, and the extension to using large beam-step sizes, which might allow higher writing throughput without loss of edge fidelity or increase in line edge roughness. TABLE I. A summary of loss improvements due to SLS for various silicon photonics features studied here. Metric Linear Waveguide Loss Curved Waveguide Loss Grating Coupler Insertion Loss Effect of Single Line Smoothing Improved by ~0.5 db/cm Improved by ~1.5 db/cm Improved by ~1.2 db Grating Coupler Uniformity Improved by ~3X Ring Resonator Q-factor Increased by ~27% 19

20 ACKNOWLEDGMENTS The authors would like to thank Uli Hofmann of GenISys GmbH for insightful discussions and non-author members of the Chrostowski research group at UBC for additional measurements. Part of this work was conducted at the Washington Nanofabrication Facility, a National Nanotechnology Coordinated Infrastructure (NNCI) site, which is supported in part by funds from the National Science Foundation. 1 R. J. Bojko, J. Li, L. He, T. Baehr-Jones, M. Hochberg, and Y. Aida, J. Vac. Sci. Technol. B 29, 06F309 (2011). 2 Y. Wang, X. Wang, J. Flueckiger, H. Yun, W. Shi, R.J. Bojko, N.A.F. Jaeger, L. Chrostowski, Opt. Express 22, (2014). 3 L. Chrostowski, M. Hochberg, Silicon Photonics Design (Cambridge University Press, Cambridge, 2015). 4 Z. Lu, J. Jhoja, J. Klein, X. Wang, A. Liu, J. Flueckiger, J. Pond, L. Chrostowski, Opt. Express 25, 9712 (2017). 20

21 Tables TABLE I. A summary of loss improvements due to SLS for various silicon photonics features studied here. Metric Linear Waveguide Loss Curved Waveguide Loss Grating Coupler Insertion Loss Effect of Single Line Smoothing Improved by ~0.5 db/cm Improved by ~1.5 db/cm Improved by ~1.2 db Grating Coupler Uniformity Improved by ~3X Ring Resonator Q-factor Increased by ~27% 21

22 Figure Captions FIG. 1. Subwavelength surface grating coupler as designed in CAD software. FIG. 2. CAD layout for grating coupler insertion loss testing featuring 2 subwavelength grating couplers spaced a fiber-width apart to accommodate testing setup connected by the shortest possible length of waveguide. Curvature is designed to not exceed critical angle above which light is more likely to scatter into the surrounding material. FIG. 3. CAD layout for waveguides showing curved and straight designs of varied lengths. Curved segments at ends of straight guides are for design compaction only and are assumed to have a negligible impact. FIG. 4. CAD layout for ring resonators showing the insertion couplers (connected) and the resonance measurement coupler (single coupler only). FIG. 5. Illustrative waveguide (not to design size) at an arbitrary angle, standard fracturing (shapes not shown), with simulated shot placements shown. Shots are not shown at full size for clarity of illustration. Note that shot placements on the edges approximate the actual shape edge, but with significant alternations that contribute to edge roughness. 22

23 FIG. 6. Illustrative waveguide (not to design size) at an arbitrary angle, SLS fracturing (shapes not shown), with simulated shot placements shown. Shots are not shown at full size for clarity of illustration. Note that shot placements on the edges approximate the actual shape edge more closely than the standard fracturing shot placement by placing each shot on the 1 nm machine grid instead of the 6 nm shot pitch. FIG. 7. (Color online) Example of data collected from automated measurements and displayed as a line graph of transmission (db) versus wavelength (nm). Examples are shown are characteristic for (a) grating coupler, (b) waveguide, and (c) ring resonator test structures. The multiple lines in (b) represent the different lengths of waveguides being measured. FIG. 8. (Color Online) Box and whisker plot showing grating coupler insertion loss values for both the baseline and SLS fracturing methods. FIG. 9. (Color Online) Line graphs showing transmission (db) versus design length (µm) for both linear and curved waveguides for both baseline and SLS fracturing methods across both production runs studied. FIG. 10. (Color online) Box and whisker plot showing ring resonator total Q-factor values for both the baseline and SLS fracturing methods. 23