Photonic Integrated Waveguide
A strand of sculpted silicon thinner than a virus, guiding light the way a river guides water.
A photonic integrated waveguide confines light inside a high-refractive-index core — silicon or silicon nitride — surrounded by a lower-index cladding of silicon dioxide, using total internal reflection to guide the light around bends and through complex optical circuits. The waveguide is the fundamental building block of a photonic integrated circuit (PIC): other functions — modulators, splitters, filters, ring resonators, and photodetectors — are all created by shaping the waveguide geometry or placing active materials next to it. Entire optical systems that once occupied a benchtop of discrete components can be collapsed onto a chip the size of a postage stamp, manufactured using the same lithography tools as silicon electronics.
Light is confined by total internal reflection between a high-index core (silicon at n ≈ 3.5, silicon nitride at n ≈ 2.0) and a lower-index silicon dioxide cladding (n ≈ 1.45). For a standard silicon-on-insulator waveguide, the core is a rectangle roughly 220 nm tall by 450–500 nm wide, patterned by deep-ultraviolet immersion lithography at 193 nm wavelength — the same tooling used for advanced CMOS transistors. At this sub-micron scale, only one or a few spatial modes of light fit inside the core, producing a single-mode waveguide that avoids the modal dispersion of multimode fibres. Propagation loss arises primarily from nanometre-scale roughness on the etched sidewalls, which scatters light out of the guided mode; state-of-the-art silicon waveguides achieve 0.28–1.55 dB/cm at 1550 nm, while silicon nitride waveguides reach 0.165 dB/cm on optimised PECVD or LPCVD processes, and best-in-class anneal-free SiN resonators have demonstrated 1.77 dB/m (0.0177 dB/cm) for thin dilute-mode waveguides. Every time light must enter or exit the chip — through a grating coupler or edge coupler — it loses another 1–3 dB per interface, a loss budget that cascades and limits the maximum circuit depth.
In plain terms
You know how a garden hose keeps water from spraying everywhere and sends it exactly where you point it? A photonic waveguide does the same thing to light. The 'hose' is a tiny ridge of silicon only a few hundred nanometres wide — about 200 times thinner than a human hair — and the light bounces along the inside because the ridge has a higher refractive index than the glass surrounding it, just like how a glass fibre works, but shrunk to chip scale.
Also called: PIC waveguide, silicon photonic waveguide, SiN waveguide, integrated optical waveguide.
Why it matters
- Collapses optical systems onto semiconductor wafers at CMOS-compatible feature sizes, enabling production of millions of identical optical chips using existing fab infrastructure — imec's 300 mm wafer platform and AIM Photonics' Albany fab demonstrate this at scale.
- Silicon's large refractive index contrast (Δn ≈ 2 versus silica) allows waveguide bends with radii of 1–5 μm, enabling optical circuits 1000× denser than silica planar lightwave circuits and 10,000× denser than discrete fibre optics.
- Silicon nitride waveguides extend the transparency window from 400 nm to 2500 nm and offer much lower nonlinear loss than silicon, enabling frequency comb generation, broadband coherent LIDAR, and low-noise microwave photonics.
- Co-packaged optics for data centre switches — integrating PIC waveguides directly with switching ASICs — is the near-term route to breaking the power and bandwidth bottleneck in hyperscale data centres, with Lightmatter demonstrating 1.6 Tbps throughput per fibre on their Passage CPO chiplet in early 2026.
- Loss budget honesty: silicon waveguide propagation loss (0.3–3 dB/cm), grating coupler insertion loss (~3–4 dB), and routing in even a modest circuit accumulate to tens of dB, fundamentally limiting the number of optical operations that can be cascaded before the signal is too weak to detect.
Where it lives
- imec 300 mm wafer-scale SiN photonic platform (Leuven, Belgium) — demonstrated 0.165 dB/cm LPCVD SiN propagation loss and 56 cm cross-wafer waveguides with <0.01 dB stitching loss per interface, used for co-packaged optics research.
- AIM Photonics Multi-Project Wafer (MPW) service (Albany NanoTech Complex, NY) — open-access 300 mm silicon photonics foundry offering active (modulators + Ge photodetectors) and passive (ultra-low-loss SiN) PIC runs to universities and companies.
- Lightmatter Passage Co-Packaged Optics chiplet — silicon photonic interconnect fabric fabricated at GlobalFoundries, demonstrating 1.6 Tbps per fibre in early 2026 sampling; used as the high-bandwidth-density interconnect between AI accelerator tiles.
- PsiQuantum photonic quantum computing chips — high-quality single-photon-compatible silicon photonic waveguides and switches designed for fault-tolerant measurement-based quantum computation; component quality has been demonstrated publicly but a full quantum processor has not been shown as of April 2026.
- NIST optical frequency comb sources — silicon nitride ring resonators and waveguides (in collaboration with university groups) used to generate octave-spanning frequency combs for optical clock comparisons and absolute frequency measurements.
- Anneal-free SiN resonators (Nature Light: Science & Applications, 2024) — research-grade platform achieving 1.77 dB/m loss and Q > 15 million at 250 °C maximum processing temperature, enabling heterogeneous integration with III–V and organic materials.
Key numbers
- Si waveguide propagation loss (state-of-the-art, 1550 nm): 0.28 – 1.55 dB/cm (0.28 dB/cm for 2 μm multimode spiral (Opto-Electronic Advances, 2023); 1.55 dB/cm for optimised 500 nm strip waveguide (IEEE JLT); standard PDK waveguides often 2–3 dB/cm)
- SiN waveguide propagation loss (LPCVD, imec): 0.165 dB/cm (Measured on 300 mm wafer-scale platform with <0.01 dB stitching loss; best-in-class thin-mode SiN reaches 0.034 dB/m = 0.00034 dB/cm (Nature, 2024))
- Fibre-to-chip insertion loss (grating coupler): 2 – 4 dB per facet (Edge couplers on imec hybrid Si+SiN platform: ~1.5 dB/fibre; grating couplers typically 3–4 dB; each interface is a fixed loss independent of circuit complexity)
- Waveguide core dimensions (standard SOI): 220 nm × 450–500 nm (Set by single-mode condition at 1550 nm telecom wavelength on 220 nm SOI; bending radii down to 1–5 μm achievable)
- Lightmatter Passage CPO throughput (2026 sampling): 1.6 Tbps per fibre (Record demonstrated by Lightmatter in March 2026 on co-packaged optics chiplet fabricated at GlobalFoundries)
Surprising facts
- A standard silicon photonic waveguide is 220 nm tall — roughly 40 times smaller than the wavelength of the light it guides. The light's electric field extends well outside the core into the surrounding glass cladding, which is why sidewall roughness only a few nanometres deep can scatter a significant fraction of the light.
- Silicon nitride waveguides processed with LPCVD and high-temperature annealing have demonstrated propagation losses of 0.034 dB/m — so low that a beam of light could travel over 130 metres through the waveguide before losing half its power, despite the chip fitting in a shirt pocket.
- Every time light must jump from an optical fibre into a silicon photonic chip through a grating coupler, it typically loses 3–4 dB — meaning 50–60 % of the photons never make it onto the chip at all. This single interface loss often exceeds the total propagation loss of the entire on-chip circuit.
- Photonic circuits are not yet remotely close to transistor counts in electronics: a state-of-the-art silicon photonic chip in 2025 contains hundreds to a few thousand optical components, while a contemporaneous CMOS processor has tens of billions of transistors. The discrepancy is driven by the much larger minimum feature size set by the wavelength of light versus the de Broglie wavelength of electrons.
- A silicon photonic ring resonator — a waveguide looped into a circle a few tens of micrometres across — can store a single wavelength channel of light so efficiently that it can completely switch a laser beam on or off with a voltage swing of less than one volt and an energy cost below one femtojoule per bit.
Open questions
- Light leaks out of a waveguide even when it is perfectly straight: Even a flawless, perfectly straight photonic waveguide with zero sidewall roughness would still lose light because the guided mode's evanescent tail penetrates the cladding and can scatter off any atomic-scale structural disorder in the glass. More striking: in ultra-high-Q SiN ring resonators, the dominant loss mechanism shifts from sidewall roughness (which engineers have largely tamed) to absorption by residual O–H and N–H bond overtones in the silica cladding — a material that was considered 'lossless' in bulk fibre optics. The 2024 anneal-free SiN work in Nature Light: Science & Applications showed that deuterated oxide cladding (replacing Si–H with Si–D shifts the absorption overtone out of the telecom window) was as important as the nitride core quality for reaching sub-2 dB/m losses.
- A photonic ring resonator can generate dozens of new laser colours from a single pump: A silicon nitride ring resonator pumped by a single continuous-wave laser can spontaneously generate a comb of hundreds of equally spaced optical frequencies spanning an octave or more — from near-infrared to mid-infrared — through the Kerr effect (four-wave mixing). This microcomb requires no population inversion, no gain medium, and no external modulation: the nonlinearity of the waveguide material alone drives the parametric frequency conversion. The 2023 Nature Photonics demonstration showed that coupling a small auxiliary ring to the main resonator to shift just the pump mode's resonance frequency boosted power conversion efficiency to over 50 %, overturning the long-held expectation that anomalous-dispersion soliton microcombs must operate at low conversion efficiency.
- PsiQuantum's loss tolerance numbers are theoretical, not system-demonstrated: PsiQuantum's 2025 arXiv study on fusion-based quantum computing claimed that advanced adaptive schemes can raise the photon loss-per-photon threshold (LPPT) to 15–18.8 %, which would make photonic quantum computing feasible even with realistic chip-scale losses. This is a counterintuitive result — most people assume any photon loss is catastrophic for quantum information. However, as of April 2026 the result comes exclusively from company-authored simulation, has not been independently reproduced experimentally, and the LPPT values assume idealised designs that have not been demonstrated at system scale. The loss budget from waveguides, grating couplers, switches, and chip-to-chip interconnects in a full million-qubit system has not been measured. The theoretical direction is credible; the specific engineering path from here to there remains undemonstrated.
A short history
The concept of guiding light in a planar waveguide was proposed by Kao and Hockham (and independently by Miller at Bell Labs) in the 1960s, building on the total internal reflection principle known since Tyndall's 1854 experiments. Silicon-on-insulator waveguides were first studied at AT&T Bell Labs in the late 1980s, but the required sub-micron lithography was not available at scale until the mid-2000s, when deep-UV immersion scanners developed for 90 nm CMOS became accessible. The landmark demonstration that silicon waveguides could modulate light at gigabit rates (Intel, 2004) and that germanium detectors could be integrated on silicon (MIT, 2006) defined the modern silicon photonics roadmap. The AIM Photonics institute in the US and imec in Belgium opened multi-project wafer services in the 2010s that democratised access to the technology. Silicon nitride emerged as the low-loss complement to silicon after Kippenberg's group at EPFL demonstrated microresonator frequency combs in SiN around 2010.
Sources
- Morin et al., 'Anneal-free ultra-low loss silicon nitride integrated photonics,' Nature Light: Science & Applications 13, 174 (2024) (https://www.nature.com/articles/s41377-024-01503-4)
- imec, 'Interfacing silicon photonics for high-density co-packaged optics,' imec technical article (November 2024) (https://www.imec-int.com/en/articles/interfacing-silicon-photonics-high-density-co-packaged-optics)
- Fahrenkopf et al., 'The AIM Photonics MPW: A Highly Accessible Cutting Edge Technology for Integrated Photonics,' IEEE J. Sel. Top. Quantum Electron. 25, 8200410 (2019) (https://www.analogphotonics.com/wp-content/uploads/2021/08/7_NFahrenkopf_JSTQE_2019.pdf)
- Lightmatter, 'Lightmatter Achieves Record 1.6 Tbps Per Fiber,' press release (March 2026) (https://lightmatter.co/press-release/lightmatter-achieves-record-1-6-tbps-per-fiber-to-accelerate-ai-optical-interconnect/)
- arXiv:2506.17999, 'Analysis of Photonic Circuit Losses with Machine Learning in Silicon Nitride Integrated Photonics' (June 2025) (https://arxiv.org/html/2506.17999)