Edge Coupler vs Grating Coupler: Choosing Fiber I/O for a Photonic Chip
The decision every PIC project makes once: edge coupling or grating coupling. Loss, bandwidth, polarization, wafer-level test, packaging, and the hybrid strategy most real programs land on.
Scope
Getting light between fiber and chip is the first problem every photonic integrated circuit ships with, and there are exactly two mainstream answers: through the facet (edge coupler) or through the surface (grating coupler). This page is the decision framework. It covers the numbers, the second-order consequences, and the hybrid pattern most production programs converge on. The hands-on alignment procedures live in the companion articles: edge coupler alignment and grating coupler first light.
The decision table
| Criterion | Edge coupler | Grating coupler |
|---|---|---|
| Insertion loss (typ / best) | 1–3 dB / under 0.5 dB with SSC + lensed or high-NA fiber | 2–4 dB / ~1 dB with optimized, sometimes bottom-mirrored designs |
| Optical bandwidth (1 dB) | Over 100 nm; often the full band and beyond | 30–40 nm; one band, chosen at design time |
| Polarization | Both TE and TM couple | Single polarization by nature (2D gratings buy diversity at ~1 dB extra) |
| Access location | Chip perimeter only, after dicing/facet prep | Anywhere on the die surface, on full wafers |
| Wafer-level test | No, needs facets | Yes, the decisive advantage |
| Alignment tolerance (1 dB) | ~±0.5–1 µm; tight | ~±2 µm; forgiving |
| Fiber geometry | In-plane butt/lensed; low-profile packages | Near-vertical (8–12° typical) or bent-fiber/mirror tricks |
| Downstream extras | None; broadband and pol-agnostic by construction | Band and polarization already filtered |
What the table actually means
Bandwidth and polarization are the physics divide. A grating diffracts: its coupling condition is wavelength- and polarization-selective by construction. An edge coupler is an adiabatic mode transformer. Nothing about it prefers a wavelength. Any application spanning O+C bands, running dual polarization (coherent receivers with their PSR front ends notwithstanding), or carrying broadband/comb light effectively has its answer already: edge.
Wafer-level test is the economics divide. A grating anywhere on the surface means every die gets optically screened before dicing: known-good-die flows, process monitoring, statistics across the wafer. Edge-only designs learn which dies work after singulation and facet prep, when most of the cost is already spent. This single row explains why gratings persist despite losing most head-to-head optical comparisons.
Packaging pulls opposite directions. Edge coupling wants polished facets, sub-micron active alignment, and fiber arrays butted in-plane (v-groove blocks are mature, and the tight tolerance is automated away in volume). Grating coupling wants near-vertical fiber, which complicates flat/low-profile modules but simplifies temporary probing enormously: a fiber on a positioner over the surface, no facet prep, ten seconds to light.
The pattern real programs use
The standard answer is both: grating couplers as sacrificial test taps on every wafer (loopbacks, process-control structures, per-die screening) and edge couplers as the product I/O that ships. The gratings die at dicing or sit unused in the corner of the die; the edge couplers carry traffic. Datacom transceiver silicon largely follows this; research chips often stay grating-only (probe convenience wins when nothing ships); LiDAR, sensing, and coherent products with bandwidth or polarization needs go edge-first and accept the test-flow cost.
Numbers worth re-verifying on your own platform rather than trusting from any table (this one included): the foundry PDK's coupler loss is specified for a particular fiber (lensed spot size or high-NA MFD) and a particular facet recipe. Swap either and the number moves dB-scale. The mode-mismatch arithmetic in the alignment article shows how fast.
Related: inverse taper and spot-size converter (the structures inside an edge coupler), lensed fiber, mode field diameter, and the Q-extraction procedure whose test flow is usually the first customer of whichever coupler you chose.