Chapter 15Grating Couplers
A grating etched into the waveguide diffracts light up towards a fibre, so a chip can be tested anywhere on the wafer.
- write the phase-matching condition and draw its wavevector diagram;
- relate period, angle and wavelength, and choose a period;
- explain directionality and the role of the buried-oxide thickness;
- estimate coupling efficiency from directionality and mode overlap, and the bandwidth from the fibre’s angular acceptance;
- explain back-reflection fringes, and describe apodised, focusing, SWG and 2D grating couplers.
Which period and angle couple 1550 nm TE light into a fibre tilted at 8°, and how much of the light can you expect to get?
15.1Phase matching
A silicon waveguide’s mode is 0.5 µm wide; a single-mode fibre’s is 10 µm wide and arrives from above. A grating coupler bridges the two in one step. The waveguide is first widened to about 12 µm, so that the mode is as wide as the fibre’s, and then runs into a shallow periodic etch. Each tooth scatters a little light, and the scattered waves add up in one direction, the direction in which their phases agree.
Take the grating along \(z\) with period \(\Lambda\), and let the guided mode in the grating region have effective index \(n_\mathrm{eff}\). The periodicity can add or remove whole multiples of the grating vector \(K = 2\pi/\Lambda\) to the mode’s propagation constant. A wave leaves the grating at angle \(\theta\) to the normal into a cladding of index \(n_c\) when its component along \(z\) matches, \(k_0n_c\sin\theta = \beta - K\), for the first diffraction order. Dividing by \(k_0 = 2\pi/\lambda\):
The second relation is Snell’s law at the flat top of the cladding: whatever the cladding, the condition can be written with the fibre angle in air. The wavevector diagram draws the same thing: an arrow of length \(n_\mathrm{eff}\) along the waveguide, a step \(\lambda/\Lambda\) back, and the result must land inside the circle of radius \(n_c\) for the light to escape upwards. It lands inside the oxide circle below just as well, so some light always goes down into the buried oxide (Section 15.2). The grating’s \(n_\mathrm{eff}\) lies between those of the unetched and etched slabs (Chapter 5); for a first estimate it is their average, weighted by the fill factor (the unetched fraction of each period).
In 220 nm SOI with a 70 nm etch, the TE slab indices at 1550 nm are 2.848 (220 nm) and 2.539 (150 nm), so a 50 % fill factor gives \(n_\mathrm{eff} \approx 2.694\). With the fibre tilted at 8° in air, \(\sin 8° = 0.139\) and
\[\Lambda = \frac{\lambda}{n_\mathrm{eff} - \sin\theta} = \frac{1.55}{2.694 - 0.139} = 0.607\ \mu\mathrm{m}.\]The averaged index is only an estimate: a full simulation of the same grating puts the period nearer 630 nm (Lab 6). In an oxide cladding the beam leaves the grating at only 5.5°, which Snell’s law turns into 8° in air.
Why not couple straight up, at \(\theta = 0\)? Then \(\Lambda = \lambda/n_\mathrm{eff}\), which is also the condition for a second-order Bragg reflection (Chapter 13): the grating reflects strongly back into the waveguide as well as up. Tilting the fibre by 8–12° moves the coupler away from that condition, and it is why fibre arrays for photonic chips are polished at an angle (Chapter 17).
Try it. Change the period, fill factor and fibre angle; the side view shows the grating to scale with the fibre at its angle.
15.2Directionality
A grating in the middle of a symmetric oxide stack scatters roughly as much light down as up. The downward light crosses the buried oxide (BOX) and meets the silicon substrate, which reflects about 17 % of it (\(r = (3.476 - 1.444)/(3.476 + 1.444) = 0.41\)). The reflected light comes back up through the grating and interferes with the light emitted upwards. Whether it adds or subtracts depends on the round-trip phase through the BOX, so the directionality, the fraction of the emitted power that goes up, oscillates with the BOX thickness \(t\) with period
\[\Delta t = \frac{\lambda}{2n_\mathrm{ox}\cos\theta_\mathrm{ox}} = \frac{1.55}{2 \times 1.444 \times 0.995} = 0.54\ \mu\mathrm{m}.\]A simple model treats the grating as a thin sheet that emits equally up and down above a reflector with field reflection \(r\). The upward field is then \(1 + re^{j\phi}\) times the emitted field, while \(1 - r^2\) of the downward power is lost in the substrate. The directionality swings between 29 % and 71 % with the substrate’s reflection. In this model the best BOX thicknesses are 1.89 and 2.43 µm, and the common 2 µm BOX gives 62 %. A real grating does not emit symmetrically, because its etch profile is asymmetric, so the true optimum is found by simulation. But the lesson stands: the BOX thickness is part of the grating-coupler design, and a coupler designed for one wafer stack may lose a decibel on another.
Designs that push the directionality higher:
- a thicker grating: a polysilicon overlay on the teeth, or a dual-etch profile, makes the grating emit more upward than downward;
- a bottom reflector: a metal layer or a Bragg stack under the BOX returns 90 % or more of the downward light; efficiencies around 1 dB or better have been reported this way [66];
- a top reflector for couplers read from below through the substrate.
15.3Mode matching to the fibre
Even with perfect directionality the light must also fill the fibre’s mode. In a uniform grating each tooth scatters the same fraction of the light still in the waveguide, so the power left decays exponentially, \(P(z) = e^{-2\alpha z}\), and the beam that leaves has the same decaying profile. Here \(\alpha\) is the grating strength, the field decay rate, set by the etch depth and the fill factor. The coupling into the fibre is the overlap of this emitted field with the fibre’s Gaussian mode (Chapter 3):
with \(D\) the directionality and \(g(z)\) the fibre’s normalised mode field projected on the grating.
No exponential matches a Gaussian well. Optimising the strength and the fibre position gives at best about 80 % overlap (−0.97 dB) for a fibre mode of waist \(w_0 = 5.2\) µm. That needs \(1/\alpha \approx 1.46\,w_0 = 7.6\) µm, with the fibre about 3.8 µm from the start of the grating [65]. A weaker grating spreads the light over more than the fibre’s mode and lets light pass the end unused; a stronger one emits a spot that is too small.
Apodisation removes this limit. Making the grating weak at the start and stronger along its length, by varying the fill factor or the etch, shapes the emitted beam into a Gaussian. The local strength needed to emit a target profile \(G(z)\) is \(\alpha(z) = G(z)/(2[1 - \int_0^z G\,\mathrm{d}z'])\). Overlaps above 95 % follow, limited by the weakest and the strongest grating that can be made.
Try it. Change the grating strength, the fibre position and the BOX thickness, then apodise the grating or put a mirror under the BOX. The side view shows the teeth deepen along an apodised grating, and the light going up and down.
With the thin-sheet model of the demo (Table 15.1), a uniform shallow-etched grating on a 2 µm BOX loses 2.0 dB to directionality and 0.9 dB to mode mismatch: −3.0 dB in all. Moving to the 1.89 µm optimum of the BOX recovers 0.5 dB. Apodising alone fixes the overlap but not the directionality (−2.2 dB); apodising with a bottom mirror brings the model to −0.2 dB. Measured couplers fall short of these numbers through the transverse overlap, fabrication errors and back-reflection. The SiEPIC EBeam PDK documents about 4.1 dB loss and a 30.6 nm 1 dB bandwidth for one of its fully etched TE couplers [15].
Table 15.1. Efficiency budget of a TE grating coupler at 1550 nm (thin-sheet model of the demo, 20 µm grating, SMF-28). Losses in dB; transverse overlap and back-reflection not included.
| Design | Directionality | Emitted before the end | Mode overlap | Total |
|---|---|---|---|---|
| Uniform, 2 µm BOX | 62 % (−2.0) | 99.5 % (−0.02) | 81 % (−0.91) | 50 % (−3.0) |
| Uniform, 1.89 µm BOX | 71 % (−1.5) | 99.5 % (−0.02) | 81 % (−0.91) | 57 % (−2.4) |
| Apodised, 2 µm BOX | 62 % (−2.0) | 98 % (−0.08) | 99 % (−0.03) | 61 % (−2.2) |
| Apodised, bottom mirror | 97 % (−0.11) | 98 % (−0.08) | 99 % (−0.03) | 95 % (−0.22) |
| Uniform but too weak (α = 0.04 µm⁻¹) | 62 % (−2.0) | 80 % (−0.98) | 68 % (−1.7) | 34 % (−4.7) |
15.4Bandwidth and polarisation
Phase matching holds exactly at one wavelength for a given angle. At another wavelength the diffracted beam swings away from the fibre axis. Differentiating the phase-matching condition, including the dispersion of \(n_\mathrm{eff}\), gives
\[\frac{\mathrm{d}\theta}{\mathrm{d}\lambda} = -\frac{n_g - \sin\theta}{\lambda\cos\theta},\]with \(n_g\) the group index of the grating region. The fibre accepts a beam tilted by \(\Delta\theta\) with efficiency \(\exp[-(\pi w_0\Delta\theta/\lambda)^2]\), which falls by 1 dB at \(\Delta\theta = 2.6°\) for \(w_0 = 5.2\) µm. With \(n_g \approx 3.5\) this gives a 1 dB bandwidth of about 42 nm. Measured couplers are narrower, typically 30–40 nm, because their beam is not a perfect Gaussian and their directionality and strength also change with wavelength. The formula shows the knobs: a lower group index (as in silicon nitride or SWG gratings) or a smaller fibre mode widens the band.
Grating couplers are strongly polarisation-dependent. TE and TM modes have different effective indices, so a grating phase-matched for TE at 1550 nm couples TM at a quite different wavelength or angle, and hardly at all at 1550 nm. PDKs offer separate TE and TM couplers; in the lab, a polarisation controller or polarisation-maintaining fibre sets the input (Chapter 31); and in products, 2D gratings handle both (Section 15.6).
15.5Back-reflection
Every grating reflects a little light back into the waveguide, from the abrupt start of the teeth and from the residual second-order Bragg reflection. Two couplers on the same waveguide then form a Fabry–Pérot cavity, and a measured spectrum shows fringes spaced by \(\lambda^2/(2n_gL)\). For couplers 1 mm apart that is 0.29 nm. If each coupler reflects −20 dB, the fringes are 0.17 dB from peak to trough; at −15 dB they reach 0.55 dB, enough to blur the resonance of a ring or the edge of a filter under test. Back-reflection is worse for a laser on the chip, whose linewidth and stability suffer from feedback (Chapter 25). Couplers with engineered first teeth, curved lines and SWG teeth bring the reflection to below −20 dB [51].
15.6Focusing, SWG and 2D grating couplers
- Focusing couplers. Instead of a straight grating fed by a taper several hundred micrometres long, curve the teeth along confocal ellipses so that the grating itself focuses the light into the narrow waveguide. The coupler becomes several times shorter with no loss of efficiency [67]. Many PDK couplers are focusing, like the SiEPIC TE coupler and the Luxtelligence TFLN coupler in the 3D gallery.
- SWG couplers. The shallow etch needs a second lithography step. Replacing it with a full etch whose teeth are sub-wavelength gratings (Chapter 13) sets the effective index by the fill factor and makes the coupler with the single etch that defines the waveguides. This is ideal for rapid prototyping by electron-beam lithography [51].
- 2D couplers. A grating periodic in two directions couples the two polarisations of the fibre into two orthogonal waveguides, both in TE: a polarisation splitter and coupler in one (Chapter 14) [65].
- Other platforms. Silicon nitride’s lower index contrast gives weaker, longer, more broadband gratings (the CORNERSTONE SiN elliptical coupler in the 3D gallery); thin-film lithium niobate couplers are similar in principle, with the crystal’s birefringence to handle.
15.7Grating couplers in practice
Grating couplers make wafer-scale testing possible: a probe station with a fibre array can reach any die without dicing (Chapter 31). They need no facet preparation, they tolerate a few micrometres of misalignment (Chapter 17), and they can sit anywhere on the chip. The price is a few decibels of loss, a bandwidth of a few tens of nanometres and a single polarisation. Edge couplers, the subject of the next chapter, make the opposite trade.
Use the PDK’s focusing TE grating coupler at 1550 nm, placed on the 127 µm pitch of the fibre array, with all couplers facing the same way. Add a loopback, two couplers joined by a short waveguide. Its spectrum gives the coupler’s peak efficiency and bandwidth, and it is subtracted from every other measurement (Chapter 32).
Reading
- [65] Dirk Taillaert, Frederik Van Laere, Melanie Ayre, Wim Bogaerts et al. “Grating couplers for coupling between optical fibers and nanophotonic waveguides.” Japanese Journal of Applied Physics, 45(8A):6071–6077, 2006.
- [66] Riccardo Marchetti, Cosimo Lacava, Lee Carroll, Kamil Gradkowski and Paolo Minzioni. “Coupling strategies for silicon photonics integrated chips.” Photonics Research, 7(2):201–239, 2019. Open access.
- [67] Frederik Van Laere, Tom Claes, Jonathan Schrauwen, Stijn Scheerlinck, Wim Bogaerts, Dirk Taillaert, Liam O’Faolain, Dries Van Thourhout and Roel Baets. “Compact focusing grating couplers for silicon-on-insulator integrated circuits.” IEEE Photonics Technology Letters, 19(23):1919–1921, 2007.
- [51] Yun Wang, Xu Wang, Jonas Flueckiger, Han Yun, Wei Shi, Richard Bojko, Nicolas A. F. Jaeger and Lukas Chrostowski. “Focusing sub-wavelength grating couplers with low back reflections for rapid prototyping of silicon photonic circuits.” Optics Express, 22(17):20652–20662, 2014. Open access.
- [1] Lukas Chrostowski and Michael Hochberg. Silicon Photonics Design: From Devices to Systems. Cambridge University Press, 2015. https://www.cambridge.org/core/books/silicon-photonics-design/BF3CF13E8542BCE67FD2BBC7104ECEAB
- [15] Lukas Chrostowski and contributors. SiEPIC EBeam PDK & Library, for SiEPIC-Tools and KLayout (MIT licence). https://github.com/SiEPIC/SiEPIC_EBeam_PDK
- Phase matching, \(n_\mathrm{eff} - \sin\theta_\mathrm{air} = \lambda/\Lambda\), sets the period: about 0.61–0.63 µm for 1550 nm TE light and an 8° fibre in 220 nm SOI.
- Fibres are tilted by 8–12° to avoid the second-order Bragg reflection of vertical coupling.
- Efficiency = directionality × mode overlap. The substrate reflection makes the directionality depend on the BOX thickness, with a 0.54 µm period; a uniform grating overlaps a Gaussian by 80 % at best.
- Apodisation raises the overlap above 95 %; bottom reflectors raise the directionality above 90 %.
- The 1 dB bandwidth (30–40 nm) comes from the fibre’s angular acceptance; couplers are polarisation-dependent, and their back-reflections make fringes in measured spectra.