Open Access
Add to BibSonomy
Abstract
Waveguiding across the visible spectrum in an unmodified bulk CMOS chip is reported. The chip is fabricated in a standard CMOS process, and a simple wet etch removes metal in predetermined locations to expose glass rib waveguides. A modified Euler bend is introduced to improve bend radii by nearly an order of magnitude in the rib waveguides, and upper-bound losses are measured at visible wavelengths. These losses range from 6.2 dB/cm at 450 nm to 3.2 dB/cm at 650 nm and represent the lowest losses reported at visible wavelengths in unmodified bulk CMOS.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Despite widespread interest in the integration of photonics and electronics on the same chip, this dream has only been realized in silicon-on-insulator platforms [1]. Bulk CMOS accounts for over 90% of the CMOS market but, until recently, has lacked a comparable solution without modifications to the semiconductor manufacturing process [2]. Furthermore, the monolithic integration community has primarily emphasized communications applications, overlooking abundant opportunities in biomedicine, projection, and imaging. The potential integration of electronic processing with high-resolution imaging [3,4], fluorescent biomarker monitoring [5], optogenetics [6], and other functionalities in a single CMOS chip may allow miniaturized systems at significantly lower costs compared to existing equipment. However, these functionalities have historically revolved around the use of visible wavelengths, a range of the spectrum that is inherently incompatible with silicon-based waveguides.
Subtractive photonics solves this problem by patterning waveguides in the back-end-of-line (BEOL) layers of a standard electronics process [7]. These layers are normally used for I/O and routing between transistors, and here the metal layers are also used as sacrificial layers for creating waveguides. These BEOL layers consist of various silicon oxides that have indices and transparencies similar to silicon dioxide, and the air displacing the etched metal provides the index contrast necessary for optical confinement. A rib geometry is used for mechanical support, where a ridge sits on a slab and supporting arms extend from the slab (Fig. 1). This geometry has been demonstrated to provide waveguiding at 1550 nm and 780 nm but requires a large bend radius to avoid being overwhelmed by bend loss [2].
In this paper, glass waveguides are demonstrated in an unmodified bulk CMOS process at visible wavelengths. A modified Euler bend is introduced to improve bend radii by nearly an order of magnitude without sacrificing mechanical integrity. Test structures are fabricated in a foundry 180 nm CMOS process and measured using three different approaches to determine the losses at visible wavelengths.
2. SiO2 waveguides
The waveguides are fabricated in a standard electronics process by drawing metal polygons on the BEOL layers. These polygons are treated identically to the metal used for routing, meaning that electronics and photonics are fabricated simultaneously. The metal is treated as a negative mold for the waveguide and is selectively wet-etched in post-processing through openings incorporated in the layout. We use Aluminum Etchant Type A, which etches not only aluminum but also tungsten and copper, for 6-10 hours at 80$^{\circ }$C to expose the rib geometry.
All waveguides demonstrated here use oxides on the via 5, metal 5, and via 4 layers for the ridge and metal 4 and via 3 layers for the slab. The layers above and below the waveguide are set to be metal (air after etching) to yield index contrast in the vertical direction. The ridge width is chosen to be 3 µm to match the mode diameter of the lensed fiber used in measurement, removing the need for a taper when edge coupling. The size of the ridge results in strong optical confinement at the cost of multimode behavior. The slab width is set to 12 µm to isolate the mode, which is mostly confined within the ridge, from the periodic supports that would otherwise cause additional scattering. These supports are periodically placed every 100 µm, each extending 3.5 µm from the slab and with a width of 20 µm to provide sufficient mechanical strength while leaving openings for the etchant to remove metal under the waveguide.
To measure waveguide loss, two structures are fabricated. The first is a 6000 µm long S-shaped waveguide that provides an upper bound on the waveguide loss by absorbing fiber coupling losses. The second uses cascaded Y-junctions to yield a differential path-length structure with optoelectronic readout using CMOS photodiodes. Scanning electron microscope (SEM) images of the structures are shown in Fig. 2.
Fig. 2. SEM images of the (a) rib waveguide, (b) roughness on the horizontal and vertical faces of the ridge, (c) waveguide-to-photodiode coupler, and (d) Y-splitter.
The line running along the sidewall of the ridge in Fig. 2(b) shows the interface between the interlayer dielectrics, and we note that the sidewall is the primary contributor of roughness in the waveguide. This is expected to be the case in more advanced processes as well due to the use of chemical-mechanical polishing (CMP) and other planarization techniques that smoothen horizontal faces. The arrays of protrusions seen in Fig. 2(c) and 2(d) are due to metal dummy fill, which exists on every metal layer of a standard CMOS process to aid in planarization but can be excluded in critical areas, such as waveguides and photodiodes.
3. Modified Euler bend
The primary limitation to the density of photonic components using these glass rib waveguides is the bend radius, with size trading with loss. The optical loss in the bend can be explained through the conformal index transformation [8]. The effect of the bend on an optical mode can be captured by replacing the bend with a straight waveguide with an equivalent index that varies over the cross section (Fig. 3). The equivalent index is described by
where $r'$ is the radial distance from the center, $n(r')$ is the index profile of the cross-section, and $R$ is the radius of curvature of the bend. It is clear that there are primarily two loss mechanisms in the bend. First, there exists a mode mismatch at the interface between a straight waveguide and radial bend solely due to the change in index profile. Second, it becomes possible for confined modes to be coupled into the slab at the outer edge, even in the absence of roughness. These can be minimized by using large bends - that is, for large $R$ the equivalent index profile in the bend approaches that of a straight waveguide. However, it is possible to design a tighter bend in a way that alleviates both of these issues.Fig. 3. Index and mode profiles based on the conformal index transformation for (a) a straight rib waveguide, (b) a bent rib waveguide, and (c) a single-sided bent rib waveguide.
The mismatch at the interface of the bend and straight sections can be minimized by linearly increasing the curvature up to the apex of the bend, followed by a linear decrease back to a straight section [9]. This bend profile is known as an Euler bend, and the section of linearly increasing curvature is parametrically defined by the Fresnel integrals
where $x$ and $y$ describe the normalized coordinates along the bend and $s$ is swept to create half of the Euler bend. Mirroring is used to implement the section with linearly decreasing curvature.The issue of coupling from confined modes within the ridge to the slab can be minimized by removing the outer slab during the bend, decreasing the equivalent index of the surrounding material and allowing the bend radius to be decreased significantly while maintaining comparable insertion loss. In the process, a new trade-off of area versus loss appears at the transition from double-sided straight waveguide to single-sided bend. We have found that tapering the outer slab of a straight waveguide in anticipation of the single-sided bend allows us to maintain both a small footprint and low bend loss, as illustrated in Fig. 4(a). We choose to taper the outer slab over a 200 µm region here, but it can be decreased or even entirely removed in return for a slight increase in insertion loss.
Fig. 4. (a) Illustration of a bend with a tapered outer slab. (b) Bend loss for standard bends, Euler bends, single-sided bends, and single-sided Euler bends versus bend radius.
Combining these solutions leaves us with a compact bend consisting of tapered transitions in and out of a single-sided Euler bend. A comparison of bend loss versus bend radius for standard bends, Euler bends, single-sided bends, and single-sided Euler bends can be seen in Fig. 4(b), which demonstrates that an Euler bend with a single-sided slab provides the minimum bend loss over a large range of radii. The bend loss is computed using the eigenmode expansion technique, and the bend radius reported for Euler bends is the radius of curvature at the apex. We choose to use a radius of 80 μm in this work.
4. Visible spectrum measurements
The optical input used in measurement is generated by a supercontinuum laser, an NKT SuperK Fianium, and fed into a tunable bandpass filter, an NKT Varia. This is delivered to the CMOS chip using a lensed fiber with a 2 µm spot size at 780 nm. All measurements are taken with a filter bandwidth of 10 nm centered at the stated wavelength. The measurement setup and chip can be seen in Fig. 5.
Fig. 5. (a) Measurement setup and (b) die photo of the CMOS chip with a horizontal field width of 2 mm.
Figure 6 shows the S-shaped waveguide with scattering at selected wavelengths using a top-view camera. Due to decreased scattering at shorter wavelengths, the power at 450 nm had to be significantly increased to improve visualization. Reflections at the tip of the output lensed fiber indicate that waveguiding still occurs down to 410 nm, but accurate measurement becomes difficult so we limit characterization to 450 nm.
Fig. 6. Microphotograph of waveguide and scattering at selected visible wavelengths. The optical input is applied at the left facet, and the horizontal field width for the scattering images is 1.8 mm.
An upper bound for the waveguide loss is determined by subtracting the loss through the fiber from the insertion loss of the waveguide. Dividing this value by the length of 0.6 cm gives us the waveguide loss in dB/cm, though this is an overestimate as the fiber coupling loss is absorbed into the insertion loss. This is plotted versus wavelength in 50 nm intervals in Fig. 7 for three different chips. The differences between chips are primarily attributed to debris falling on the waveguides of chips 2 and 3, rather than differences in fabrication or etching, as the debris is visible under an optical microscope. Capped waveguides with small openings can be used to avoid this issue in the future.
The test structure containing cascaded symmetric Y-splitters with optoelectronic readout is shown in Fig. 8. Photodiode couplers with a coupling efficiency of 71% at 650 nm are used to couple light from the waveguides into standard CMOS photodiodes (Fig. 2(c) and 9) [2]. The loss of the waveguide-to-photodiode coupler and reponsivity of the photodiode are inherently de-embedded through the differential nature of the measurement. The photodiode has an active area of 9$\times$18 µm$^2$ and uses the P-well/deep N-well and deep N-well/P-substrate junctions at zero-bias, as seen in Fig. 9. Also, the splitter exhibits negligible loss in simulation, and the mirror symmetry of the splitter means it can be conservatively de-embedded by subtracting 3 dB from the loss. As a result, the waveguide loss in dB/cm can be computed as
where $I_{PD}$ is the DC photodiode current, $n$ and $m$ are indices of the photodiodes, and the difference in length propagated between successive photodiodes is 0.185 cm. Unfortunately, some of the light from the input fiber coupled into the 2nd and 4th photodiodes, which are located on the same side of the chip as the fiber. Therefore, we only consider the 1st and 3rd photodiodes, which have a difference of 7.23 dB in current and yield a waveguide loss of 3.16 dB/cm at 650 nm. We only consider 650 nm as we have increasingly worse control over the optical mode appearing at each Y-splitter as the wavelength gets shorter.Fig. 8. Microphotograph and scattering at 650 nm from the optoelectronic differential test structure.
Fig. 9. Waveguide-to-photodiode coupler (a) cross section and profile of power through the coupler at a wavelength of 650 nm, (c) efficiency, and (d) photodiode structure.
Finally, the top-view camera method [10] is used to determine the waveguide loss for the best performing chip from Fig. 7. The bandpass filter applied to the supercontinuum source is swept to provide images of scattering at multiple wavelengths, such as those seen in Fig. 6. These images are digitally processed to compare the relative intensities of scattering between the section before the first bend and at the end of the waveguide, with the latter chosen at a location of minimum scattering to avoid underestimating the loss. The propagation losses computed using this method are shown alongside the two previously mentioned methods in Fig. 10.
Fig. 10. Waveguide loss as a function of wavelength using the optical insertion loss (chip 1), top-view camera, and differential path length methods.
5. Conclusion
We report foundry-fabricated photonic waveguides in an unmodified 180 nm bulk-CMOS process. These waveguides are created using the method of subtractive photonics and are measured in the visible spectrum to present losses as low as 3.2 dB/cm at 650 nm. The multiple BEOL layers available are used to create modified Euler bends that improve bend radii by nearly an order of magnitude. The broad transparencies of the waveguides, along with their close integration with electronics, present subtractive photonics as a promising platform for monolithic bioelectrophotonic systems in CMOS.
Funding
NSF graduate fellowship (NSF Grant No. DGE‐1745301); Carver Mead New Adventures Fund; Space Solar Power Project.
Acknowledgments
This work was supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1745301. The authors also thank the following: Space Solar Power Project and Carver Mead New Adventures Fund for support, Muse Semiconductor for their tapeout services, and Craig Ives and Volkan Gurses for useful discussions.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. M. Rakowski, C. Meagher, K. Nummy, et al., “45nm cmos-silicon photonics monolithic technology (45clo) for next-generation, low power and high speed optical interconnects,” in Optical Fiber Communication Conference, (Optica Publishing Group, 2020), pp. T3H-3.
2. C. Ives, D. Sarkar, and A. Hajimiri, “Subtractive photonics in bulk cmos,” IEEE J. Solid-State Circuits 58(11), 3030–3043 (2023). [CrossRef]
3. G. Yurtsever, B. Považay, A. Alex, B. Zabihian, W. Drexler, and R. Baets, “Photonic integrated mach-zehnder interferometer with an on-chip reference arm for optical coherence tomography,” Biomed. Opt. Express 5(4), 1050–1061 (2014). [CrossRef]
4. P. Muellner, E. Melnik, G. Koppitsch, J. Kraft, F. Schrank, and R. Hainberger, “Cmos-compatible si3n4 waveguides for optical biosensing,” Procedia Eng. 120, 578–581 (2015). [CrossRef]
5. L. Hong, H. Li, H. Yang, and K. Sengupta, “Fully integrated fluorescence biosensors on-chip employing multi-functional nanoplasmonic optical structures in CMOS,” IEEE J. Solid-State Circuits 52(9), 2388–2406 (2017). [CrossRef]
6. F. Wu, E. Stark, M. Im, I.-J. Cho, E.-S. Yoon, G. Buzsáki, K. D. Wise, and E. Yoon, “An implantable neural probe with monolithically integrated dielectric waveguide and recording electrodes for optogenetics applications,” J. Neural Eng. 10(5), 056012 (2013). [CrossRef]
7. R. Fatemi, C. Ives, A. Khachaturian, and A. Hajimiri, “Subtractive photonics,” Opt. Express 29(2), 877–893 (2021). [CrossRef]
8. A. Yariv, Optical electronics in modern communications, vol. 1 (Oxford University Press, USA, 1997).
9. F. Vogelbacher, S. Nevlacsil, M. Sagmeister, J. Kraft, K. Unterrainer, and R. Hainberger, “Analysis of silicon nitride partial euler waveguide bends,” Opt. Express 27(22), 31394–31406 (2019). [CrossRef]
10. J. D. Bradley, C. C. Evans, J. T. Choy, O. Reshef, P. B. Deotare, F. Parsy, K. C. Phillips, M. Lončar, and E. Mazur, “Submicrometer-wide amorphous and polycrystalline anatase tio 2 waveguides for microphotonic devices,” Opt. Express 20(21), 23821–23831 (2012). [CrossRef]
