Fiber delay lines for array calibration: what a millimetre and a kelvin actually cost you

Started by kenji_rf · 2026-11-17 14:32 UTC · 19 views · 4 replies
K Member
Every wideband array I have helped calibrate ends up in the same argument: buy a fiber true-time-delay bank, or keep the coax and fix the phase in software. Both can work. The difference is which error survives the temperature cycle. Here is the arithmetic I run before the purchase order.

1. Delay per metre, and what a millimetre costs
- SMF at 1550 nm carries about 4.9 ns/m of group delay (n_g around 1.47).
- At 10 GHz, 1 ps of delay error is 3.6 degrees of phase. So 1 mm of length error is 4.9 ps, i.e. 17.6 degrees, which on its own breaks a plus/minus 5 degree calibration spec.
- A cable house will hand you pigtails at plus/minus 10 mm unless you ask otherwise: 176 degrees at 10 GHz. You cannot absorb that in software across a 2 GHz instantaneous band, because the error scales with frequency, not as a constant offset. You trim the fiber, then measure the residual per channel on a VNA and carry it as a table.
- Trim target: plus/minus 0.1 mm, which is plus/minus 0.5 ps and 1.8 degrees at 10 GHz. A fusion splicer with a clean cleave holds that; a mechanical splice does not, and it also moves the reflection point.

2. Thermal drift is 40 ps/km/K and it is mostly common-mode
- Silica delay drifts about 40 ps per km per K at 1550 nm, roughly 7 ppm/K, of which about 5 ppm is index and 2 ppm is expansion.
- On a 100 m spool that is 4 ps/K, or 14 degrees/K at 10 GHz. A rack that swings 10 K across a day moves 140 degrees. That is where the "it calibrated fine in the lab" reports come from.
- The useful part: it is common-mode. Put every spool in one insulated enclosure and the differential drift is set by how evenly that enclosure heats, not by the fiber itself.
- The failure I have actually seen: spools mounted at different heights in a 19 inch box with a 3 K top-to-bottom gradient. Two channels 30 cm apart drifted 12 ps relative to each other, which is 43 degrees at 10 GHz, and it looked exactly like a bad module.

3. Group delay ripple is the spec that bites, not average delay
- A reflection in the pigtail creates two-path interference. With 490 ns of delay (100 m), the ripple period is 1/(2 tau), about 1 MHz. That sits inside your channel, so it is invisible on a CW phase measurement and obvious as group delay ripple on a VNA sweep.
- Consequences for the spec: require group delay ripple under 1 ps across the band, return loss above 55 dB at the connectors and above 45 dB at the splice, and use APC everywhere rather than UPC-with-adapter hygiene.
- If you only buy "delay plus/minus 1 ps" and never look at ripple, you will pass incoming inspection and fail the range test.

4. Where the coax case still wins
- At 10 GHz, 10 m of good coax costs 3 to 5 dB, its phase tracks temperature far worse per unit length than fiber, and every re-mate adds a new phase offset you have to re-measure.
- Coax still wins on two counts: no bias loop drifting underneath the delay, and a phase trimmer you can turn with a screwdriver at 2 am. A fiber bank needs a stable optical source and a defined operating point before its delay means anything.

5. Acceptance sequence I use
a) Per-channel group delay at 25 C from the phase slope across the band, three points minimum.
b) Soak the enclosure at 45 C for 4 hours, re-measure, and fit a per-channel coefficient instead of assuming the nominal 40 ps/km/K.
c) VNA group delay sweep, looking specifically for periodicity under 10 MHz, which is the reflection signature.
d) Only then write the calibration table, with the temperature coefficient as its own column so the field team can correct for rack temperature instead of re-calibrating.

Discussion: what differential drift do you specify across channels, and do you correct it with a temperature table or by holding the enclosure at a fixed setpoint?

Disclosure: our production RFoF modules come from F-tone Networks (f-tone.com). The numbers above are from our own incoming inspection and range work, not from their datasheet.
S Member
Three additions from the same set of failures, all earned in the field rather than on the bench.

On the differential drift spec: we write plus/minus 1.5 ps channel-to-channel over the enclosure's own temperature range, and we verify it in a gradient rig instead of a soak. A 4-hour soak at 45 C only proves the setpoint. The case you describe - spools at different heights in a 19 inch box - is a spatial gradient, so the test that catches it is two reference spools, one at the top and one at the bottom of the rack, with their relative delay tracked while the box is heated from one side. That finds the 12 ps in twenty minutes instead of after a range test.

On trimming to plus/minus 0.1 mm: measure the group delay at the operating wavelength, not at 1310 nm. With dispersion around 17 ps/nm/km, trimming 1550 nm hardware against a 1310 nm measurement leaves roughly 2.5 ps per 100 m, which is half your budget before anything warms up. Same argument for the pigtail: a vendor who quotes delay at 1310 for a 1550 link is telling you the length, not the delay.

On the ripple spec, make it peak-to-peak over any 100 MHz window rather than "under 1 ps across the band". A single reflection gives you a sinusoid, and its peak sits wherever the two-path delay puts it; a whole-band number lets a bad unit hide between the peaks. Ask the vendor for the S21 phase sweep as a file, and check the reflection periodicity yourself - a 1 MHz ripple is the signature of a pigtail reflection, and nothing else in the chain produces it.

Two smaller items. Require return loss at the operating wavelength and at the temperature extremes, because a connector that holds 55 dB at 25 C can sit at 48 dB at 60 C once the ferrule gap opens. And when the bank sits downstream of the modulator, put a 1 percent tap and log optical power next to the delay data: a bias loop creeping 0.5 percent gives you a gain step that looks exactly like a delay step on a single-tone measurement.

The screwdriver argument for coax still holds at 2 am. What changed my mind was the re-mate cost: every connector mate on a calibrated channel adds a phase offset that has to go back into the table, and on a 64 channel array that is the expensive part, not the cable.
L Admin
Thank you, kenji_rf — carrying delay per millimetre, the 40 ps/km/K common-mode term, and group-delay ripple from pigtail reflections on the same traveler is the right way to stop a "it calibrated in the lab" surprise. Also noted SI_Practice_Han's follow-up.

Editorial keepers (soft promo = 0):
1. At 10 GHz, 1 mm is already ~17.6 degrees; a ±10 mm pigtail tolerance cannot be absorbed as a constant software offset across a wide instantaneous band. Trim, then keep a per-channel residual table.
2. Put every spool in one enclosure and treat a top-to-bottom rack gradient as a differential-delay source, not a module fault. Fit a measured per-channel coefficient after a warm soak rather than assuming the nominal 40 ps/km/K.
3. Require group-delay ripple (look for a ~1/(2τ) period from a pigtail reflection), not only average delay. APC and a return-loss floor belong next to the ripple spec.

Open: what differential drift do you specify across channels, and do you correct it with a temperature table or by holding the enclosure at a setpoint?

— rfforum 编辑部 / Editorial Team
M Member
Marking this - our project starts something similar next month, learning from everyone's experience.
F Member
The per-metre arithmetic is right, and the number I would add is the thermal coefficient of the delay itself: about 40 ppm per degree C for standard SMF, so a 30 m spool swinging 40 C moves roughly 6 ps, which is 21 degrees at 10 GHz before you touch the length tolerance. That is why the software phase fix keeps winning on cost even when the fiber bank is technically cleaner. What we ended up doing was keeping the coax for the reference channel only and putting the delay bank where the temperature is held to plus/minus 1 C, which cut the daily recalibration from four times to once.
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.