Fiber Array Attach to Photonic Chips: V-Groove Arrays, Alignment Tolerances, Epoxy, and the Loss Budget

Scope

This article covers the permanent attachment of a multi-channel fiber array unit (FAU) to the edge or surface of a photonic integrated circuit: what the array is and which of its tolerances matter, how much misalignment each coupler type forgives, how the adhesive is chosen and cured, what loss to expect per channel, and how the finished joint is qualified. Single-fiber pigtailing of lasers and detectors, lensed-fiber probing during test (Active fiber alignment to edge couplers), and the choice between coupler types (Edge coupler versus grating coupler) have their own pages. Laser-welded and solder-attached arrays are mentioned only where they differ from epoxy.

The fiber array unit

A fiber array unit is a row of stripped fibers laid in etched or precision-ground V-grooves in a glass or silicon substrate, capped with a lid, bonded, and polished as one block. The block is what gets aligned and glued; the fibers never move relative to each other again. The numbers that matter are the ones that set how well the row of cores matches the row of couplers on the chip.

Pitch is 127 µm or 250 µm as standard (127 µm for most PIC work; 250 µm matches ribbon fiber). Vendors quote the V-groove spacing to ±0.5 µm and the resulting core position to ±0.7 µm for arrays up to 16 channels, relaxing to ±1.0 µm at 48 channels and ±1.5 µm at 72 (Corning OEM-039). Typical core placement on a good array is 0.5 µm. Part of that budget is the fiber itself: SMF-28 Ultra is specified at 125.0 ± 0.7 µm cladding diameter and ≤ 0.5 µm core-to-cladding concentricity, and its mode field diameter at 1550 nm is 10.4 ± 0.5 µm (Corning PI1424). Because the groove locates the cladding and the light is in the core, the concentricity error goes straight into the channel-to-channel alignment and cannot be tuned out; only the array as a whole is aligned. Channel counts run from one to 96 in standard product; the array can also carry polarization-maintaining or small-core (UHNA) fibers, at a slightly wider position tolerance.

The end face is polished flat (flatness about 1.6 µm typical) at 0° or at 8° ± 0.3° for edge coupling, and at a steeper custom angle for surface coupling. A 0° array against a chip facet returns 14–20 dB; anything above about 5° of polish pushes the return loss past 50 dB (Corning OEM-039), which is why 8° is the default. For grating couplers the two common geometries are an array polished so that the fiber meets the chip surface at the grating's design angle (Tyndall's packaging rules standardize on 7° incidence) and a total-internal-reflection block polished near 40°, in which the light turns inside the fiber and leaves through its side, so the array lies flat on the chip (Li 2014; Carroll 2016).

Tolerance arithmetic

For two Gaussian modes of 1/e² radii $w_1$ and $w_2$ displaced laterally by $d$, the coupling efficiency is

$$\eta \;=\; \left(\frac{2 w_1 w_2}{w_1^2 + w_2^2}\right)^{2} \exp\!\left(-\frac{2 d^2}{w_1^2 + w_2^2}\right)$$

with the first factor the mode-size mismatch and the second the offset (Marcuse 1977). For equal modes the lateral offset that costs 1 dB is $d = 0.48\,w$, or 0.24 times the mode field diameter. A tilt $\theta$ between equal modes costs $\exp[-(\pi n w \theta/\lambda)^2]$, and a gap $z$ between them costs $[1 + (\lambda z / 2\pi n w^2)^2]^{-1}$, where $n$ is the index of whatever fills the gap. The Fiber Coupling Efficiency Calculator evaluates these for arbitrary inputs; the table gives the values that decide an attach process at 1550 nm:

Measured values match: Avdeev and colleagues report 1 dB lateral tolerances of ± 2.8 µm for SMF-28, ± 1.4 µm for a 4 µm MFD fiber, and ± 1.2 µm for UHNA-7 against the same edge coupler (arXiv:2405.11980, 2024), the slight excess over the Gaussian estimate coming from the coupler mode being wider than the fiber mode. The mismatch factor on its own is the reason spot-size converters exist: an SMF-28 mode butted directly to a 4 µm coupler mode loses 3.5 dB before any misalignment, and to a 3 µm mode 5.5 dB; the same fiber against a 6 µm mode loses 1.25 dB.

Grating couplers trade the other way. Taillaert and colleagues measured under 0.5 dB penalty for 1 µm of lateral offset and 0.5 dB for 2° of angle, with the peak wavelength moving about 9 nm per degree; Carroll and colleagues put the practical 1 dB window at about ± 2.5 µm for a grating and ± 0.5 µm for an edge coupler. The angular sensitivity is what makes the polish angle of a grating-coupler array a specification rather than a preference: 0.3° of polish error is 3 nm of spectral shift on every channel at once.

Two consequences follow for arrays. First, the whole row shares one alignment, so the channel-to-channel core error (0.5–0.7 µm) is spent from the same budget as the process error; an edge-coupled array with 4 µm mode couplers has about ± 1 µm to work with in total, and a 16-channel array at ± 0.7 µm core position has already used most of it. Second, pitch error accumulates: a 0.05 % pitch mismatch between array and chip (127.06 µm against 127.00 µm) is 0.9 µm of offset at the sixteenth channel. Edge-coupled arrays therefore run with the largest chip mode the design can afford, and grating-coupled arrays can accept cheaper array tolerances at the price of bandwidth.

Facet and coupler preparation

Edge coupling needs a facet that is flat, perpendicular within the tilt tolerance above, and set back from the waveguide end by a known distance. Deep-etched facets made in the same lithography as the waveguides (reactive plasma dicing gives facet angles within a degree of vertical; Avdeev 2024) beat mechanically polished facets on this, because the coupler-to-facet distance is defined by the mask rather than by the polishing stop. Any gap between array and facet is filled by the adhesive, and the gap loss in the table above should be evaluated with the adhesive index rather than air: at $n = 1.5$ the 1 dB gap for a 4 µm mode grows from 8 µm to about 12 µm.

The adhesive also handles the reflection. Silica to air is 3.4 % (14.7 dB return loss); silica to an epoxy of index 1.56 is 0.13 % (28.7 dB), and with the 8° polish the residual is directed out of the mode, so the joint as a whole reaches the 50–60 dB that a narrow-linewidth source needs (Optical feedback in semiconductor lasers explains why that number matters). A 0° array with an index-matched bond is acceptable only when the source can tolerate 30 dB return loss.

Procedure

1. Inspect the array under a microscope for chipped cores, debris in the groove region, and polish scratches; measure the fiber-to-fiber pitch on the polished face against the chip's coupler pitch if the vendor report is not available.
2. Clean the chip facet or surface with a solvent wipe and a plasma or UV-ozone clean; adhesive adhesion to the chip is set here.
3. Mount the chip on a temperature-controlled vacuum chuck and the array in a gripper on a six-axis stage. Set the array angle so that the beam leaving the polished face lands on the waveguide axis. For an 8° silica face against air the beam deviates 3.6° from the fiber axis (Snell's law with $n = 1.45$); once the gap is filled with an adhesive of index 1.55 the deviation is about 0.5°, so the final angle is set with the adhesive present in step 8. For a grating array, set the design incidence angle.
4. Bring the array to within a few tens of micrometres of the facet on camera. Launch light into an outer channel and into the chip's alignment loop if it has one; a loop-back waveguide that connects two array channels lets the whole search run on transmitted power through the array alone.
5. Search laterally in the outer channel until light couples, then optimize all six axes in turn on the loop-back power. Then optimize on the two outermost channels together: the difference between them is the pitch and yaw error, and it is minimized by rotating the array about the axis normal to the facet rather than by translating it.
6. Record the per-channel coupled power. Against the tolerance table, decide whether the spread across channels is array error (fixed) or residual yaw (correctable), and correct what can be corrected.
7. Retract by the intended bond-line thickness, dispense the adhesive so that it wets the full facet contact area and the underside of the array where it meets the chip or a submount, and return to the optimum. Adhesive is drawn into the gap by capillary action; check that no bubble sits on a channel.
8. Re-optimize with the adhesive in place (the index change moves the optimum slightly), then cure under UV without releasing the gripper. Watch the coupled power during the cure and log the shift.
9. Release the gripper, thermally post-cure per the adhesive datasheet, and re-measure every channel. The difference between the pre-cure optimum and the post-cure reading is the attach loss for the record.

An active-alignment step on a modern gantry takes on the order of a second per axis; the cycle time of the whole attach is dominated by dispensing and curing, which is why two-step cures (a fast tack cure, then the full cure off the stage) are common in production. Passive attach, in which the array is placed by machine vision against fiducials with no light on, is limited by the placement tool: general microelectronic die bonders place to about ± 10 µm, single-mode work needs ± 1–2 µm, and the tools built for it reach 0.2–0.3 µm at 3σ (Barwicz 2018; vendor data), enough for grating couplers and for edge couplers with mechanical stops.

Choosing and curing the adhesive

The adhesive has four jobs: hold the array against shock and pull, fill the optical gap with an index close to silica, survive the qualification temperatures, and move as little as possible while doing so. The datasheet numbers that map onto those jobs:

Three things to read off the table. Every one of these indices is above silica's 1.44–1.45, so the residual reflection is small but not zero, and an 8° polish is still wanted for return-loss-critical channels. The CTE roughly triples above $T_g$, so an adhesive with $T_g$ inside the 85 °C damp-heat or thermal-cycling ceiling will creep during qualification; a $T_g$ above the highest test temperature, in practice at least 95 °C for Telcordia programs, is the usual rule. And shrinkage acts on the bond line: 1.5 % linear shrinkage across a 20 µm bond line is 0.3 µm of motion, which is a third of the 1 dB budget for a 4 µm coupler, whereas 0.39 % is 0.08 µm. Thin bond lines, low-shrink formulations, and symmetric adhesive placement (so the shrink pulls along the optical axis rather than across it) are how the cure shift is kept to a few tenths of a dB. Published attach studies report total shifts of that order: Wang and colleagues measured a 0.05 dB change on curing and 0.15 dB after the full process on a grating-coupled array (Photonics 12, 545, 2025).

For the same reason the array is usually bonded on two faces, the polished face to the chip facet and its underside to the chip or a submount, so that the joint is not a single cantilevered glue line. Laser-welded ferrules and solder-attached arrays remove the polymer from the optical path entirely and are used where hermeticity or high-temperature operation requires it, at the cost of the fixturing described in pigtailing.

The loss budget per channel

A finished edge-coupled channel carries: the coupler's own mode-mismatch loss at zero offset (0.2–1 dB for a well-designed inverse-taper or edge coupler against a matched fiber, several dB against SMF-28 without a spot-size converter), the residual alignment error after cure (0.1–0.5 dB), the gap and reflection terms (under 0.1 dB with an index-matched bond), and channel-to-channel variation from the array's core-position error (a few tenths of a dB across a 16-channel row at the tolerances above). Reported results sit where that arithmetic says they should: 1.1–1.5 dB per facet with SMF-28 and ± 0.2 dB uniformity across an eight-channel array, in that case attached by CO₂-laser fusion rather than epoxy (Nauriyal 2023), 1.3 dB from patch cable to chip in IBM's automated attach (Barwicz 2018), and about 1.5 dB for SMF-28 against 0.15 dB for a matched small-core fiber on the same coupler (Avdeev 2024). Grating-coupled arrays add the packaging excess over the probe-station value; Li and colleagues measured 1.2 dB of excess on a 6.5 dB fiber-to-fiber link with the 40° block.

The numbers are worth stating per channel and per facet, because the array attach is where a chip's characterized insertion loss (Coupler loss de-embedding) becomes the module's guaranteed loss. A 16-channel transceiver with two attached arrays and a 0.5 dB variation across each array has a 1 dB spread in link budget between its best and worst lane before anything electronic is counted.

Qualification

Telcordia GR-1221 (passive optical components) is the reference most packaging programs test against, with GR-468 for active modules. The conditions vendors report running: thermal cycling from −40 to +85 °C for 500 cycles; damp heat at 85 °C and 85 % relative humidity for 2000 h; high- and low-temperature storage at 85 °C and −40 °C for 2000 h each; mechanical shock of 500 g for 1 ms; vibration at 20 g from 20 to 2000 Hz; fiber retention at 0.45–0.5 kgf held for one minute, three times; and side pull at 0.23–0.45 kgf at 90°. The usual pass criterion is a change in insertion loss of no more than 0.5 dB, with 0.2 dB imposed by stricter programs. Published results for well-made epoxy joints are comfortably inside that: Kumagai and colleagues report −0.11 to +0.11 dB after 2000 h of damp heat and ± 0.07 dB after 500 thermal cycles (SEI Technical Review 89, 2019).

Two of these tests target the adhesive directly. Thermal cycling exercises the CTE mismatch between array (silica or silicon), adhesive (55–200 ppm/K), and chip (silicon, about 3 ppm/K): over the 125 K span of the test a 20 µm bond line of a 74 ppm/K adhesive changes thickness by 0.19 µm, and it is the shear at the interfaces, not the thickness, that opens cracks. Damp heat exercises moisture uptake and the drop in $T_g$ that comes with it. Both are why the adhesive's $T_g$ and CTE are chosen before its index.

Common failure modes

A channel-to-channel loss ramp across the array that cannot be nulled by yaw is a pitch mismatch between array and chip; measure the array pitch and re-order or re-specify. A uniform loss step after cure is shrinkage along the wrong axis or an asymmetric fillet; move the dispense point. A single bad channel on an otherwise good array is usually debris or a bubble in the gap, visible from the side under a microscope before the cure. A loss that drifts by tenths of a dB over the first days is an under-cured adhesive still relaxing; extend the post-cure. A return loss that measures 20 dB instead of 50 dB on an 8° array means the array was mounted with the polish angle in the wrong plane, or the gap was not filled. And a facet that couples well on the bench but poorly after attach often has a polished facet whose coupler set-back was never measured; the adhesive-filled gap is then longer than the design assumed, which the gap term in the table above turns into loss at the smallest mode sizes first.

Packaging is where much of a photonic module's cost sits; Tyndall's IPIC puts it at as much as 80 % of product cost in some applications. The array attach is the largest single item in that fraction, and the numbers above are the ones that decide whether it is done once.

References: Corning, "Fiber Array Units," product information OEM-039-AEN (2019). Corning, "SMF-28 Ultra Optical Fiber," product information PI1424 (2025). OZ Optics, "V-Groove Assemblies," DTS0083. SQS Vláknová optika, fiber array unit specifications. D. Marcuse, "Loss analysis of single-mode fiber splices," Bell System Technical Journal 56, 703 (1977). L. Carroll et al., "Photonic packaging: transforming silicon photonic integrated circuits into photonic devices," Applied Sciences 6, 426 (2016). D. Taillaert et al., "Grating couplers for coupling between optical fibers and nanophotonic waveguides," Japanese Journal of Applied Physics 45, 6071 (2006). C. Li et al., "Silicon photonics packaging with lateral fiber coupling to apodized grating coupler embedded circuit," Optics Express 22, 24235 (2014). R. Marchetti et al., "Coupling strategies for silicon photonics integrated chips," Photonics Research 7, 201 (2019). S. Avdeev et al., arXiv:2405.11980 (2024). T. Barwicz et al., "Automated, high-throughput photonic packaging," Optical Fiber Technology 44, 24 (2018). J. Nauriyal, M. Song, Y. Zhang, M. Granados-Baez, and J. Cardenas, "Fiber array to chip attach using laser fusion splicing for low loss," Optics Express 31, 21863 (2023). Norland Products, NOA 61 and NOA 68 technical data sheets. Epoxy Technology, OG142-112, OG198-54 and 353ND technical data sheets. Dymax, OP-4-20632 product data sheet. Telcordia GR-1221-CORE, Generic Reliability Assurance Requirements for Passive Optical Components; GR-468-CORE for optoelectronic devices. IEC 61300-2-4, Fibre retention. K. Kumagai et al., SEI Technical Review 89 (2019). Y. Wang et al., Photonics 12, 545 (2025). Tyndall National Institute, Irish Photonic Integration Centre, packaging and hybrid integration overview. Europractice, Photonic packaging design rules v1.7 (2024).