MFD Matching: Cutting Fiber-to-Chip Coupling Loss in Silicon Photonics
A standard single-mode fiber butted against a silicon photonic waveguide loses about 5.5 dB before alignment error is even considered. The cause is mode-field mismatch, and the fix is a different fiber — used only where it is needed.
If you are coupling a standard single-mode fiber straight onto a silicon photonic waveguide and losing several decibels at the interface, the alignment stage is probably not the problem. The dominant loss term is usually a mismatch between two mode fields that were never the same size to begin with.
Where the loss actually comes from
Standard telecom fiber such as SMF-28 carries a mode field diameter of roughly 10.4 µm at 1550 nm. A silicon photonic edge coupler, even with an inverse-taper spot-size converter, typically presents something in the 3 µm range. Those two modes overlap poorly, and the penalty follows directly from the overlap integral of two Gaussian beams:
Put real numbers in and the scale of the problem is obvious. These are pure mode-mismatch figures — perfect alignment, no Fresnel reflection, no angular or lateral offset:
| Fiber MFD | Chip MFD | Coupling efficiency | Mismatch loss |
|---|---|---|---|
| 10.4 µm (SMF-28) | 3.0 µm | 0.28 | 5.5 dB |
| 10.4 µm (SMF-28) | 4.0 µm | 0.45 | 3.5 dB |
| 6.0 µm | 4.0 µm | 0.85 | 0.7 dB |
| 4.0 µm | 3.0 µm | 0.92 | 0.35 dB |
| 3.3 µm | 3.0 µm | 0.99 | 0.04 dB |
Five and a half decibels lost before anything else goes wrong. No amount of active alignment recovers that, because the light is not misplaced — it is the wrong size. And note how sharply the curve flattens: closing the gap from 10.4/3.0 to 4.0/3.0 removes more than 5 dB, while the last 0.7 µm of matching is worth only a few tenths.
Matching the mode instead of fighting it
The fix is to change the fiber, not the alignment. Ultra-high-numerical-aperture (UHNA) fiber has a much smaller core and correspondingly smaller mode field — small enough to sit in the same range as the chip. Splice it to the array end and the mismatch term largely disappears.
The obvious question is why the whole assembly is not simply built from UHNA fiber. Two reasons: UHNA has higher attenuation than conventional single-mode fiber, and it does not mate cleanly with the standard connectors and equipment on the rest of the link. So it is used only where it is needed.
What the construction actually looks like
The drawing below is a 12-channel MFD fiber array on 250 µm pitch. The detail worth noticing is the length of UHNA fiber: about 10 mm, running from the polished end face to a splice point just outside the substrate. Past that splice it is ordinary Corning fiber all the way to the FC/APC connectors, 1 m away.
So the part is a hybrid by design: a short mode-matched section where the optics demand it, and a conventional, connectorizable fiber everywhere else. The V-groove geometry itself is unchanged from a standard array — 60° groove angle, Ø125 µm cladding, 250 µm pitch, 0.375 mm edge margin.
The splice is not free
Every design choice costs something, and here the cost sits at the splice between two very different fibers. Two consequences are worth planning around.
Polarization extinction ratio drops. A straight PM fiber array is specified at PER ≥ 20 dB. The PM version of the MFD array is specified at ≥ 16 dB, typically ≥ 18 dB. That difference is the splice: joining UHNA to PM panda fiber means aligning two birefringent axes across a junction, and the alignment is never perfect. If your system budget assumes 20 dB PER, check it against 16 before committing.
Insertion loss is specified per version. The PM configuration is specified at IL ≤ 0.5 dB (typically 0.3 dB); the SM configuration at ≤ 0.8 dB. Both figures already include the splice contribution.
Choosing the MFD
Standard MFD values are 6.0, 7.5 and 9.0 µm at 1310/1550 nm, and the range is customizable from 3.3 to 10 µm (down to 2.6 µm on the SM version). MFD tolerance is ±0.5 µm, typically ±0.3 µm.
Two practical points when specifying:
- Match to the chip, not to the smallest number available. The table above shows the payoff flattening quickly. Once you are within about 1 µm of the chip mode, further reduction buys tenths of a decibel while making the assembly harder.
- Give a target MFD, not a fiber part number. UHNA1, 3, 4 and 7 are all available; which one lands closest to your waveguide depends on the wavelength and the taper design. Tell us the mode you need to hit.
Where it fits
MFD-matched arrays are specified where the fiber-to-chip interface is the limiting loss in the link: silicon photonics packaging, PIC edge coupling, coherent receivers, optical engines, and quantum photonic setups where launch efficiency into a small mode determines whether the experiment closes at all.
Channel counts run 1 to 32 on the PM version and up to 48 on the SM version, at 127 or 250 µm pitch. Everything else — polish angle, fiber length, connector, substrate outline — is built to your drawing.