Chapter 10Y-Branches and Multimode Interference Couplers

Splitting light evenly sounds trivial; doing it over a wide band with tiny loss takes some care.

In this chapter you will learn to
  • explain the Y-branch, and why a combiner fed from one arm passes only half the power;
  • compute which modes an abrupt junction excites, and explain the snaking of an offset input;
  • derive the self-imaging lengths of a multimode waveguide;
  • design 1 × 2 and 2 × 2 MMIs and estimate their bandwidth and fabrication tolerance;
  • choose between directional couplers, Y-branches, MMIs and adiabatic couplers.
Design question

Y-branch or MMI, and how long must the MMI be?

10.1Y-branches

A Y-branch splits one waveguide into two by widening it gradually and separating it into two arms. If the transition is adiabatic (Chapter 8), the fundamental mode of the input evolves into the even supermode of the two arms, which carries half the power in each: a 50/50 splitter, symmetric by construction and therefore broadband.

Run backwards, the same device is a combiner, and here it holds a surprise. Feed light into only one arm and it is the sum of an even and an odd supermode of the arms, in equal parts. Only the even supermode evolves into the output’s fundamental mode; the odd one has nowhere to go in a single-mode output and radiates away. A symmetric Y-combiner fed from one arm passes exactly half the power, and the same is true for two inputs out of phase. This is not a defect but the Y-junction’s reciprocity, and it is why a Mach–Zehnder interferometer built from Y-branches loses its light into the substrate rather than into a second output port at its nulls (Chapter 11).

Practical Y-branches face a fabrication limit: the gap between the arms cannot close to zero, and the blunt tip scatters. The compact Y-junction of the SiEPIC EBeam PDK solved this by optimising the shape of a short taper with particle-swarm optimisation in 3D FDTD, reaching an insertion loss of about 0.28 dB [46, 15]. You can turn it over in the 3D gallery.

10.2Mode excitation at an abrupt junction

When a narrow waveguide meets a wide one abruptly, its mode \(E_\mathrm{in}\) is projected onto the modes \(\phi_\nu\) of the wide section, with amplitudes given by overlap integrals (Chapter 8):

\[c_\nu = \frac{\int E_\mathrm{in}\,\phi_\nu^*\,\mathrm{d}x}{\int\lvert\phi_\nu\rvert^2\,\mathrm{d}x}.\]

Symmetry decides a lot. A centred input is even, so it has zero overlap with every odd mode (\(\nu = 1, 3, 5, \dots\)) and excites only even ones. An input offset from the centre excites both; the even and odd modes then beat against each other as they travel, and the intensity snakes from side to side across the wide guide. The pattern is fixed in space, not travelling: it is the interference of modes with different propagation constants.

10.3Self-imaging in a multimode waveguide

The multimode section of width \(W\) supports many modes. In the effective-index picture it is a slab of index \(n_r\) (the vertical slab index of the film, Chapter 6) in a cladding \(n_c\), and its modes are close to sines across an effective width that includes the penetration of the field into the cladding,

\[W_e = W + \frac{\lambda_0}{\pi}\left(\frac{n_c}{n_r}\right)^{2\sigma}\left(n_r^2 - n_c^2\right)^{-1/2}, \qquad \sigma = \begin{cases} 0 & \text{field parallel to the side walls,} \\ 1 & \text{field normal to the side walls.} \end{cases}\]

The TE-like modes of an SOI waveguide have their electric field normal to the side walls: in the effective-index method of Chapter 6 their lateral slab is TM. So they take \(\sigma = 1\), and the field reaches only \((n_c/n_r)^2 \approx 0.26\) times as far into the oxide as a field parallel to the walls would.

The propagation constants of these modes are spaced quadratically, \(\beta_0 - \beta_\nu \approx \nu(\nu + 2)\pi/(3L_\pi)\), with the beat length of the two lowest modes

MMI beat length
\[L_\pi = \frac{\pi}{\beta_0 - \beta_1} \approx \frac{4n_rW_e^2}{3\lambda_0}\]

Because \(\nu(\nu + 2)\) is always either even or odd times a fixed step, all the modes come back into a definite phase relation at regular lengths, and the input field is reproduced: self-imaging [42]. Three regimes follow from which modes the input excites:

  • General interference (any input): a mirrored single image at \(3L_\pi\), a direct one at \(6L_\pi\), and \(N\)-fold images at \(3L_\pi/N\) in between, such as two images at \(3L_\pi/2\), the basis of the 2 × 2 MMI.
  • Paired interference (input at \(\pm W_e/6\), where some modes are not excited): images at \(L_\pi\), two-fold at \(L_\pi/2\); shorter 2 × 2 couplers.
  • Symmetric interference (centred input, only even modes): \(N\)-fold images at \(3L_\pi/(4N)\), four times shorter; a 1 × 2 splitter needs only \(3L_\pi/8\).

Try it. Change the MMI’s width, length and input offset; the top view shows the device to scale, and the map shows the images form.

// try it · MMI self-imaging (2D effective-index model)
geometry: top view, width stretched
intensity in the MMI (top view, length squeezed)
intensity across the output end
guided modes –beat length Lπ –1 × 2 length –power in upper / lower half –

10.4MMI design

Worked Example 10.1 · A 1 × 2 MMI in 220 nm silicon

Take a 6 µm wide multimode section in 220 nm silicon at 1550 nm (TE), with \(n_r = 2.85\) (the slab index of Chapter 5) and \(n_c = 1.444\). The field is normal to the side walls, so \(\sigma = 1\) and the effective width is \(W_e = 6 + (1.55/\pi)(1.444/2.85)^2/\sqrt{2.85^2 - 1.444^2} = 6 + 0.052 = 6.052\) µm, so

\[L_\pi = \frac{4 \times 2.85 \times 6.052^2}{3 \times 1.55} = 89.8\ \mu\mathrm{m}.\]

A centred input gives two images at \(3L_\pi/8 = 33.7\) µm, at \(\pm W_e/4 = \pm1.51\) µm from the centre, where the two output waveguides start. A 2 × 2 MMI by general interference needs \(3L_\pi/2 = 135\) µm; by paired interference, with inputs at \(\pm W_e/6\), \(L_\pi/2 = 45\) µm. Taking \(\sigma = 0\) by mistake would give \(W_e = 6.20\) µm and a 1 × 2 MMI 35.4 µm long, 1.7 µm too long, which costs about 1 dB (Lab 3).

Three practical points shape a real MMI:

  • Tapered access waveguides. Widening the input and output waveguides to 1–2 µm where they meet the multimode section makes the images sharper and the device less sensitive to the exact length, and reduces the loss from the higher-order modes that the ideal theory neglects.
  • Tolerance. \(L_\pi\) grows with \(W_e^2\), so a 1 % width error moves the images by 2 % of the length, about 0.7 µm in Example 10.1; the image region is roughly that long, so a few tens of nanometres of width error are tolerable.
  • Bandwidth. \(L_\pi\) is inversely proportional to wavelength, so the images drift as the wavelength changes; short MMIs (small \(W\), symmetric interference) are the most broadband. Excess losses of a few tenths of a dB and imbalances of a few per cent over tens of nanometres are typical of good designs.

10.5Comparing splitters

Table 10.1. Choosing a 3 dB splitter in silicon photonics.

Device Ports Bandwidth Strengths Weaknesses
Directional coupler (Ch 9) 2 × 2 narrow any ratio, compact, low loss ratio drifts with wavelength and gap
Y-branch 1 × 2 broad 50/50 by symmetry, small tip limited by lithography; one-arm combining loses 3 dB
MMI 1 × N, N × N moderate to broad robust to fabrication, any N longer, small excess loss and imbalance
Adiabatic coupler 2 × 2 broad ratio set by symmetry, very broadband long (hundreds of µm)
What this means for your design

For a 1 × 2 split on your first chip, the PDK’s Y-branch is the safe choice: small, broadband and measured. For a 2 × 2 coupler that must work across the four channels of the reference link, an MMI or an adiabatic coupler beats a directional coupler, whose split ratio drifts by a few per cent per nanometre (Chapter 9).

Explore further

Free simulations by others that pair well with this chapter [11]:

Reading

  • [42] Lucas B. Soldano and Erik C. M. Pennings. “Optical multi-mode interference devices based on self-imaging: principles and applications.” Journal of Lightwave Technology, 13(4):615–627, 1995.
  • [46] Yi Zhang, Shuyu Yang, Andy Eu-Jin Lim, Guo-Qiang Lo, Christophe Galland, Tom Baehr-Jones and Michael Hochberg. “A compact and low loss Y-junction for submicron silicon waveguide.” Optics Express, 21(1):1310–1316, 2013. 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
  • [3] Katsunari Okamoto. Fundamentals of Optical Waveguides, 3rd edition. Academic Press, 2021.
Key takeaways
  • An adiabatic Y-branch splits 50/50 by symmetry; as a combiner fed from one arm it passes only half the power.
  • An abrupt junction excites the wide section’s modes by overlap; centred inputs excite only even modes.
  • Self-imaging: \(L_\pi \approx 4n_rW_e^2/(3\lambda_0)\); images at \(3L_\pi\) (general), \(L_\pi\) (paired) and \(3L_\pi/(4N)\) (symmetric).
  • A 6 µm silicon MMI has \(L_\pi = 90\) µm (TE, \(\sigma = 1\)): a 1 × 2 splitter is 34 µm long, a 2 × 2 coupler 45–135 µm.
  • Use Y-branches for compact 1 × 2 splits, MMIs or adiabatic couplers for broadband 2 × 2 splits.