Silicon Photonics and CMOS Integration

Silicon photonics takes one result from guided-wave electromagnetism — that a high-index core buried in a lower-index cladding traps light by total internal reflection — and drives it to a material extreme the fibre never reaches. A standard fibre's core and cladding differ in index by less than a percent; in silicon-on-insulator the silicon core ( $n_{\mathrm{Si}}\approx3.48$ at $1550\,\mathrm{nm}$ ) sits in silicon dioxide ( $n_{\mathrm{SiO_2}}\approx1.44$ ), an index step of order two. That single change of magnitude is the whole subject: it drops the critical angle to about $24°$ , shrinks the guided mode to a submicron rectangle, and lets a wire of light turn a right-angle corner in a few micrometres — so an entire optical circuit fits in the footprint of a transistor block. The platform also inherits, from solid-state physics, two awkward facts about the silicon crystal. Its bandgap is indirect, so radiative recombination is hopelessly inefficient and silicon cannot practically lase — gain must be imported by bonding or growing III–V material. Its lattice is centrosymmetric, so the second-order susceptibility vanishes and there is no fast Pockels effect; modulation instead exploits the free-carrier plasma-dispersion effect in a diode threaded through the mode, and detection is handed to epitaxial germanium. The payoff is economic: photonic functions ride the trillion-dollar CMOS base, with wafer-scale reproducibility and monolithic co-integration with transistors. The costs are three loss channels — sidewall scattering, two-photon absorption, and a fierce thermo-optic drift — and managing them while staying CMOS-compatible defines the frontier.

Key takeaways

Silicon photonics guides light in submicron silicon-on-insulator (SOI) waveguides using the huge index contrast between silicon ( $n\approx3.48$ ) and silica ( $n\approx1.44$ ) — fibre physics at an index step of order two.
The contrast gives $\mathrm{NA}\approx3.17$ and $\theta_c\approx24.5°$ , so a single-mode strip waveguide is only about $500\,\mathrm{nm}\times220\,\mathrm{nm}$ and can bend within a few micrometres.
Silicon is transparent across the telecom bands ( $\hbar\omega<E_g\approx1.12\,\mathrm{eV}$ , so $\lambda>1.1\,\mu\mathrm{m}$ ) but its indirect gap makes it a poor emitter — lasers are heterogeneously integrated III–V.
Modulation uses the plasma-dispersion effect (free carriers shift $n$ and add absorption) since centrosymmetric silicon has no Pockels effect; figure of merit $V_\pi L\sim1$ – $3\,\mathrm{V\cdot cm}$ .
Loss of $1$ – $3\,\mathrm{dB/cm}$ is sidewall scattering, two-photon absorption caps on-chip power, and resonances drift at ≈ $70$ – $80\,\mathrm{pm/K}$ from silicon's $\mathrm{d}n/\mathrm{d}T\approx1.8\times10^{-4}\,\mathrm{K^{-1}}$ .
Key idea: because $\Delta n\approx2$ between silicon and its oxide, a single-mode waveguide is only a few hundred nanometres across and bends within microns — an entire optical circuit fits in the area of a transistor block and is built with the same lithography, letting photons and electrons share one chip.

The core insight

Because $\Delta n\approx2$ between silicon and its oxide cladding, a single-mode waveguide is only a few hundred nanometres across and can bend within microns — so an entire optical circuit fits in the area of a transistor block, and it can be built with the very same lithography, letting photons and electrons share one chip.

Where this lives today

The commercial engine of silicon photonics is the data centre. Pluggable and co-packaged transceivers on SOI now move data between switch chips at 400 and 800 Gb/s, integrating Mach–Zehnder or microring modulators, germanium photodetectors, and heterogeneously bonded III–V lasers millimetres from the switch ASIC — cutting energy per bit and multiplying capacity by wavelength-division multiplexing many channels down one waveguide. Foundries such as GlobalFoundries, Tower, and the AIM Photonics and IMEC lines now offer silicon-photonic process design kits the way they offer CMOS ones, so a designer draws components from a qualified library and ships a mask set. The same blocks reach past communications. Solid-state LiDAR steers a beam with no moving parts by phasing a chip-scale optical array; functionalised microrings become label-free biosensors. And the platform's low-loss, wafer-scale reproducibility has made it the leading bet for large-scale photonic quantum computing — PsiQuantum is fabricating millions of interferometers and single-photon components on standard silicon lines, wagering that the manufacturing base Soref pointed to is what a million-qubit machine will need.

The physical picture

Picture a ray bouncing inside a silicon core. Because the step to the oxide is so large, the critical angle is small — near $24°$ from the normal — so almost every ray is perfectly reflected, with only a short evanescent tail in the cladding. That short tail is the whole story of dense integration: the mode is squeezed into a rectangle a few hundred nanometres on a side, two such wires run side by side without coupling, and a silicon wire turns a right-angle corner in five micrometres where a fibre would need centimetre bend radii.

Figure

Cross-section of a single-mode SOI strip waveguide: a 500 nm × 220 nm crystalline-silicon core (n = 3.48) on buried oxide (n = 1.44) squeezes the quasi-TE mode into a submicron footprint with a short evanescent tail — beside it, a standard single-mode fibre mode of about 10 µm drawn to the same scale shows why silicon integrates so densely

Two facts about the crystal shape everything else. First, silicon cannot make its own light: its indirect bandgap means an electron and hole cannot recombine into a photon without a phonon to conserve momentum — so gain is imported, a III–V die bonded on top with its mode coupled into the silicon. Second, silicon cannot switch light by any fast, low-loss mechanism: inversion symmetry forces $\chi^{(2)}$ to zero and kills the Pockels effect. The workaround is the plasma-dispersion modulator — a capacitor whose dielectric is the waveguide: change the carrier population in a p–n or p–i–n junction threaded through the mode, and you change the local index, the phase, and, through an interferometer or resonator, the amplitude the receiver reads. Every resonant silicon device then lives on a thermal knife-edge: a fraction of a degree walks a microring off by a full linewidth. The second figure shows both faces of a microring — its comb of resonances and the drift that threatens them.

Figure

Transmission spectrum of an add-drop microring (radius ≈ 5 µm, group index ≈ 4.2): sharp resonant notches recur every free spectral range FSR ≈ 18 nm, each with linewidth ≈ 78 pm at loaded Q ≈ 2×10⁴ — the red dashed notch shows the thermo-optic drift of ≈ 80 pm per kelvin, roughly one full linewidth per degree, which is why silicon rings need active thermal control

The core equations

The platform starts from one number: the critical angle. Snell's law at the total-internal-reflection threshold reads $n_{\mathrm{Si}}\sin\theta_c=n_{\mathrm{SiO_2}}\sin90°$ , and the huge index ratio drives $\theta_c$ down and the confinement up.

\theta_c=\arcsin\!\left(\frac{n_{\mathrm{SiO_2}}}{n_{\mathrm{Si}}}\right)=\arcsin\!\left(\frac{1.44}{3.48}\right)\approx24.5° \qquad (1)

where: $n_{\mathrm{Si}}\approx3.48$ — index of the silicon core at $1550\,\mathrm{nm}$ ; $n_{\mathrm{SiO_2}}\approx1.44$ — index of the buried-oxide cladding; $\theta_c$ — critical angle from the normal. Physical reading: any ray hitting the wall at more than $24.5°$ from the normal is perfectly guided, so nearly all of the mode's angular content is trapped — hence the short tail and small mode. How wide can the core be and still carry one mode? The normalized frequency (V-number) collects geometry and contrast into one dimensionless group; single-mode operation demands $V$ below the first higher-order cutoff. Because the numerical aperture is enormous for SOI, the core must be tiny.

V=\frac{2\pi a}{\lambda}\sqrt{n_{\mathrm{core}}^2-n_{\mathrm{clad}}^2}=\frac{2\pi a}{\lambda}\,\mathrm{NA},\qquad \mathrm{NA}=\sqrt{3.48^2-1.44^2}\approx3.17 \qquad (2)

where: $a$ — half-width of the guiding core; $\lambda$ — free-space wavelength; $\mathrm{NA}$ — numerical aperture; $V$ — normalized frequency. Physical reading: with $\mathrm{NA}\approx3.17$ , twenty times that of fibre, the half-width $a$ keeping $V$ small is a few hundred nanometres — the exact rectangular cutoff, found numerically, lands near $500\,\mathrm{nm}\times220\,\mathrm{nm}$ at $1550\,\mathrm{nm}$ . To modulate, silicon leans on free carriers. An injected electron–hole gas adds a negative Drude term to the permittivity, $\varepsilon(\omega)=\varepsilon_\infty-Ne^2/[\varepsilon_0 m^*(\omega^2+i\omega/\tau)]$ ; at optical frequencies $\omega\tau\gg1$ , so expanding $n=\sqrt{\varepsilon}$ to first order gives the plasma-dispersion index shift.

\Delta n=-\frac{e^2\lambda^2}{8\pi^2 c^2\varepsilon_0 n}\left(\frac{\Delta N_e}{m_e^*}+\frac{\Delta N_h}{m_h^*}\right) \qquad (3)

where: $\Delta N_e,\Delta N_h$ — changes in electron and hole density; $m_e^*,m_h^*$ — their effective masses; $n$ — background silicon index; $e,\varepsilon_0$ — electron charge and vacuum permittivity. Physical reading: adding carriers lowers the index (Soref and Bennett fit $\Delta n$ and $\Delta\alpha$ to data) — the workhorse behind every silicon modulator, and it inescapably drags free-carrier absorption $\Delta\alpha$ along. An index change becomes a signal through phase: a shifter of length $L$ accumulates $\Delta\phi=(2\pi/\lambda)\,\Delta n_{\mathrm{eff}}L$ , and a full $\pi$ switch fixes the trade between drive voltage and length in the voltage–length product.

\Delta\phi=\frac{2\pi}{\lambda}\,\Delta n_{\mathrm{eff}}\,L,\qquad V_\pi L=\frac{\lambda}{2}\,\frac{L}{\Delta n_{\mathrm{eff}}}\bigg|_{\Delta\phi=\pi} \qquad (4)

where: $\Delta n_{\mathrm{eff}}$ — carrier-induced change in modal effective index; $L$ — active length; $V_\pi L$ — voltage–length product for a $\pi$ shift, smaller is better. Physical reading: silicon's weak, absorption-coupled $\Delta n_{\mathrm{eff}}$ forces $V_\pi L\sim1$ – $3\,\mathrm{V\cdot cm}$ , so a low-voltage modulator is millimetres long — hence compact designs wrap the shifter into a resonant microring. A microring resonates when its round trip is an integer number of guided wavelengths, and its comb repeats at the free spectral range — the resonance set by the phase index $n_{\mathrm{eff}}$ , the spacing by the group index $n_g$ .

\lambda_{\mathrm{res}}=\frac{n_{\mathrm{eff}}\,L_{\mathrm{rt}}}{m},\qquad \mathrm{FSR}=\frac{\lambda^2}{n_g\,L_{\mathrm{rt}}} \qquad (5)

where: $L_{\mathrm{rt}}$ — round-trip length; $m$ — integer mode order; $n_g=n_{\mathrm{eff}}-\lambda\,\mathrm{d}n_{\mathrm{eff}}/\mathrm{d}\lambda$ — group index. Physical reading: a smaller ring pushes resonances farther apart (larger FSR) but bends more tightly and radiates more, so range trades against Q. The same resonance that makes the ring useful makes it fragile: silicon's index rises with temperature, so the resonance red-shifts, and the drift per kelvin is large.

\frac{\mathrm{d}\lambda_{\mathrm{res}}}{\mathrm{d}T}=\lambda_{\mathrm{res}}\,\frac{1}{n_g}\frac{\mathrm{d}n_{\mathrm{eff}}}{\mathrm{d}T}\approx70\text{–}80\ \mathrm{pm/K} \qquad (6)

where: $\mathrm{d}n_{\mathrm{eff}}/\mathrm{d}T$ — thermo-optic coefficient of the mode, dominated by silicon's bulk $\mathrm{d}n/\mathrm{d}T\approx1.8\times10^{-4}\,\mathrm{K^{-1}}$ . Physical reading: with a linewidth of only tens of picometres, this drift walks the device off resonance within a single kelvin — hence integrated heaters or athermal cladding on nearly every resonant silicon chip. The tight confinement itself follows from the symmetric-slab eigenvalue $\tan(\kappa a)=\gamma/\kappa$ with $\kappa^2+\gamma^2=k_0^2(n_1^2-n_2^2)$ , recalled from aq-optics-waveguide-modes-wave-equation rather than re-derived. The constraint circle's radius is fixed purely by the contrast: for Si/SiO₂, $n_1^2-n_2^2\approx10$ against $\sim0.02$ for a fibre, so the cladding decay constant $\gamma$ is large, the evanescent tail $e^{-\gamma|x|}$ collapses to a fraction of a wavelength, and the mode shrinks to a few hundred nanometres — the physics behind dense integration.

Derivation AThe Mach–Zehnder plasma-dispersion modulator transfer function

The goal is to derive, from two-beam interference, how a bias-controlled index change in one arm of a balanced interferometer becomes an intensity the receiver reads, and to pin down the switching condition behind the $V_\pi L$ figure of merit. We feed a 3-dB splitter into two arms, shift one arm's index by $\Delta n_{\mathrm{eff}}$ with a reverse-biased p–n junction, recombine, and take the output intensity — arriving at the $\cos^2$ transfer law and the full- $\pi$ extinction condition.

Step 1. A balanced Y- or MMI splitter divides the input amplitude equally, $E_1=E_2=\tfrac{1}{\sqrt2}E_{\mathrm{in}}$ . Why: equal splitting lets the arms interfere to true zero at the dark port, giving high extinction.

Step 2. The biased arm accumulates extra phase, so $E_1\to\tfrac{1}{\sqrt2}E_{\mathrm{in}}e^{i\phi_0}$ and $E_2\to\tfrac{1}{\sqrt2}E_{\mathrm{in}}e^{i(\phi_0+\Delta\phi)}$ with $\Delta\phi=(2\pi/\lambda)\,\Delta n_{\mathrm{eff}}L$ . Why: the carriers act only through the phase they imprint, so all the drive enters here as $\Delta\phi$ .

Step 3. Recombining and taking intensity, $E_{\mathrm{out}}=\tfrac12E_{\mathrm{in}}(1+e^{i\Delta\phi})e^{i\phi_0}$ , so $T=P_{\mathrm{out}}/P_{\mathrm{in}}=\cos^2(\Delta\phi/2)$ . Why: interference is the transducer, converting an invisible phase difference into an amplitude a photodiode reads.

Step 4. Full extinction needs $\Delta\phi=\pi$ , i.e. $\Delta n_{\mathrm{eff}}=\lambda/2L$ , fixing the length and hence $V_\pi L$ . Why: silicon's weak $\Delta n_{\mathrm{eff}}$ bites here — a small index change forces a long device or high voltage, the central compromise of silicon modulators. *Verified: standard interferometer algebra; the $\cos^2$ law is the two-beam interference result of Hecht Ch. 9, giving $V_\pi L\sim1$ – $3\,\mathrm{V\cdot cm}$ for silicon.*

Go deeper · Derivation B · 1 minHow two-photon absorption caps the power a silicon wire can carry

Derivation BHow two-photon absorption caps the power a silicon wire can carry

The goal is to show why a passive silicon waveguide, transparent to a single $1550\,\mathrm{nm}$ photon, nonetheless goes lossy as the power climbs — and why the free carriers it generates make it worse than a simple nonlinear absorber. We check the photon-energy bookkeeping, write the carrier rate equation the two-photon process feeds, solve for the steady-state carrier density, and assemble the intensity-loss equation, whose $I$ and $I^2$ terms set a hard power ceiling.

Step 1. One $1550\,\mathrm{nm}$ photon carries $\hbar\omega=hc/\lambda\approx0.80\,\mathrm{eV}<E_g=1.12\,\mathrm{eV}$ and cannot bridge the gap, but two together give $1.60\,\mathrm{eV}>E_g$ and can. Why: linear absorption is negligible yet a two-photon channel opens — the loss is intrinsically intensity-dependent.

Step 2. Each two-photon event lifts an electron–hole pair, so $\mathrm{d}N/\mathrm{d}t=\beta_{\mathrm{TPA}}I^2/(2\hbar\omega)-N/\tau_{\mathrm{eff}}$ . Why: the carriers linger for a lifetime $\tau_{\mathrm{eff}}$ and add their own free-carrier absorption on top of the two-photon loss.

Step 3. In steady state $N_{\mathrm{ss}}=\beta_{\mathrm{TPA}}\tau_{\mathrm{eff}}I^2/(2\hbar\omega)$ , so longer lifetime means a denser carrier cloud. Why: the lever designers pull — a reverse-biased p–i–n sweeps carriers out, shrinking $\tau_{\mathrm{eff}}$ and the induced loss.

Step 4. Collecting terms, $\mathrm{d}I/\mathrm{d}z=-\alpha_0 I-\beta_{\mathrm{TPA}}I^2-\sigma_{\mathrm{FCA}}N_{\mathrm{ss}}I$ , a loss growing as $I$ and $I^2$ . Why: the super-linear terms cap the power a silicon wire carries before nonlinear absorption dominates — the constraint on on-chip amplifiers and nonlinear devices. *Verified: consistent with the standard silicon nonlinear-loss model (Reed and Knights); uses $\hbar\omega=hc/\lambda\approx0.80\,\mathrm{eV}$ at $1550\,\mathrm{nm}$ and $\beta_{\mathrm{TPA}}\sim0.5$ – $0.9\,\mathrm{cm/GW}$ .*

Limiting cases

Silicon photonics is defined by material and geometric extremes; the key regimes are:

Case	Result	Meaning
High-contrast confinement	$n_{\mathrm{Si}}/n_{\mathrm{SiO_2}}\approx2.4$	Small critical angle, submicron mode, few-µm bends — the enabling geometric limit.
Telecom transparency	$\hbar\omega<E_g\approx1.12\,\mathrm{eV}\Rightarrow\lambda>1.1\,\mu\mathrm{m}$	Passive and low-loss in the O- and C-bands; below ≈1.1 µm interband absorption switches on.
No electro-optic response	$\chi^{(2)}=0$ (centrosymmetric)	No Pockels effect; fast modulation must use plasma dispersion, coupling index change to absorption.
Two-photon absorption	$\alpha_{\mathrm{eff}}=\alpha_0+\beta_{\mathrm{TPA}}I$	At 1550 nm $2\hbar\omega\approx1.6\,\mathrm{eV}>E_g$ , so TPA ( $\beta_{\mathrm{TPA}}\sim0.5$ – $0.9\,\mathrm{cm/GW}$ ) and its carriers cap power.
Thermo-optic drift	$\mathrm{d}n/\mathrm{d}T\approx1.8\times10^{-4}\,\mathrm{K^{-1}}$	Resonances tune at ≈70–80 pm/K, demanding stabilisation or athermal designs.

Worked examples

1 · The critical angle and numerical aperture of SOI (confidence builder)

Using $n_{\mathrm{Si}}=3.48$ and $n_{\mathrm{SiO_2}}=1.44$ at $1550\,\mathrm{nm}$ , find the critical angle and numerical aperture of the silicon–oxide interface, and compare with a standard single-mode fibre ( $\mathrm{NA}\approx0.14$ ).

Step 1. Critical angle from equation (1): $\theta_c=\arcsin(1.44/3.48)=\arcsin(0.414)\approx\mathbf{24.5°}$ . Rays striking the wall steeper than $24.5°$ from the normal are perfectly reflected — a very wide capture cone.

Step 2. Numerical aperture from equation (2): $\mathrm{NA}=\sqrt{3.48^2-1.44^2}=\sqrt{12.11-2.07}=\sqrt{10.04}\approx\mathbf{3.17}$ .

Step 3. Compare: $\mathrm{NA}_{\mathrm{SOI}}/\mathrm{NA}_{\mathrm{fibre}}\approx3.17/0.14\approx23$ . The SOI guide has more than twenty times the numerical aperture of fibre — the reason its mode collapses to submicron scale, and the reason fibre-to-chip coupling needs a taper or grating coupler to bridge mode sizes that differ by an order of magnitude. θ_c ≈ 24.5°, NA ≈ 3.17, about 23× the fibre value.

2 · Free spectral range and thermal budget of a microring

A silicon microring of radius $R=5\,\mu\mathrm{m}$ has group index $n_g=4.2$ at $\lambda=1550\,\mathrm{nm}$ . Find its free spectral range; if its loaded $Q=2\times10^4$ , find the resonance linewidth, and estimate the temperature change that walks it off by one linewidth given a drift of $80\,\mathrm{pm/K}$ .

Step 1. Round-trip length $L_{\mathrm{rt}}=2\pi R=31.4\,\mu\mathrm{m}$ . Free spectral range from equation (5): $\mathrm{FSR}=\lambda^2/(n_g L_{\mathrm{rt}})=(2.40\times10^{-12}\,\mathrm{m}^2)/(1.32\times10^{-4}\,\mathrm{m})\approx\mathbf{18.2\ nm}$ .

Step 2. Linewidth from $Q=\lambda/\delta\lambda$ : $\delta\lambda=\lambda/Q=1550\,\mathrm{nm}/(2\times10^4)\approx\mathbf{78\ pm}$ .

Step 3. Thermal budget: at $80\,\mathrm{pm/K}$ , one linewidth of shift takes $\Delta T=78/80\approx\mathbf{1.0\ K}$ . A single kelvin walks the ring off resonance — the quantitative "thermal knife-edge," and why it needs an integrated heater with feedback. FSR ≈ 18.2 nm, linewidth ≈ 78 pm, one linewidth per ≈ 1 K.

3 · Index change to switch a Mach–Zehnder modulator

A balanced silicon Mach–Zehnder modulator has a $2\,\mathrm{mm}$ active arm at $\lambda=1550\,\mathrm{nm}$ . What effective-index change switches it from bright to dark?

Step 1. Extinction needs a $\pi$ phase difference, so from equation (4) $\Delta n_{\mathrm{eff}}=\lambda/(2L)=(1.55\times10^{-6}\,\mathrm{m})/(4\times10^{-3}\,\mathrm{m})\approx\mathbf{3.9\times10^{-4}}$ . A few parts in $10^4$ is small, but plasma dispersion delivers it only at carrier densities near $10^{17}$ – $10^{18}\,\mathrm{cm^{-3}}$ that also add free-carrier absorption; shortening the arm for speed demands more voltage — the $V_\pi L$ trade made concrete. Δn_eff ≈ 3.9×10⁻⁴.

Physics of the portrait

Silicon photonics has an unusually clear origin: for two decades it was one man's argument that the material the electronics industry had already mastered could carry light too, if you engineered around its limitations.

Origins · 1 minRichard Soref and how this was discovered

Richard Soref

1932– · father of silicon photonics; the Soref–Bennett plasma-dispersion equations

Richard A. Soref is the physicist most responsible for the idea that silicon — not III–V compounds — should be the substrate of integrated optics. From the early 1980s, for decades at the US Air Force Research Laboratory, he argued that silicon-on-insulator offered exactly what photonics needed: a transparent, high-index core at telecom wavelengths, made by an industry already spending billions on lithography. In 1987, with Brian Bennett, he published the paper quantifying how injected free carriers change silicon's index and absorption — the Soref–Bennett relations that remain the design workhorse for every silicon modulator since. He and Lorenzo had already demonstrated low-loss silicon waveguides in the mid-1980s, when the field barely existed. Soref understood the platform's two structural defects — the indirect gap that forbids lasing and the centrosymmetry that forbids the Pockels effect — and framed the whole discipline as the art of working around them: import gain from III–V, detect with germanium, modulate with plasma dispersion. When the data-centre bandwidth crisis of the 2000s made the economics irresistible, the physics was waiting.

History

1969

— Miller at Bell Labs introduces the concept of integrated optics on a planar substrate

1986

— Soref and Lorenzo demonstrate low-loss silicon waveguides at 1.3 µm

1987

— Soref and Bennett quantify the free-carrier plasma-dispersion effect in silicon

2004

— Intel reports a silicon optical modulator with bandwidth above 1 GHz

2005

— a micrometre-scale silicon microring modulator and a continuous-wave silicon Raman laser are demonstrated

Going further

Texts. Saleh and Teich, Fundamentals of Photonics, 3rd ed., Ch. 8 and Ch. 24 · Hecht, Optics, 5th ed., Ch. 5 and Ch. 9 · G. T. Reed and A. P. Knights, Silicon Photonics: An Introduction (Wiley, 2004) · L. Chrostowski and M. Hochberg, Silicon Photonics Design (Cambridge Univ. Press, 2015) · R. A. Soref and B. R. Bennett, "Electrooptical effects in silicon", IEEE J. Quantum Electron. 23, 123 (1987) · Demtröder, Laser Spectroscopy, Vol. 1.

Provenance. A verifier should recompute: (1) the critical angle $\arcsin(1.44/3.48)\approx24.5°$ and numerical aperture $\sqrt{3.48^2-1.44^2}\approx3.17$ ; (2) the photon energy $hc/\lambda\approx0.80\,\mathrm{eV}$ at 1550 nm, confirming $2\times0.80>E_g=1.12\,\mathrm{eV}$ ; (3) the slab eigenvalue $\tan(\kappa a)=\gamma/\kappa$ and MZI transfer $T=\cos^2(\Delta\phi/2)$ ; (4) the ring free spectral range $\lambda^2/(n_g L_{\mathrm{rt}})\approx18.2\,\mathrm{nm}$ for $R=5\,\mu\mathrm{m}$ , $n_g=4.2$ . Material constants (Si 3.48, SiO₂ 1.44, $E_g=1.12\,\mathrm{eV}$ , $\mathrm{d}n/\mathrm{d}T\approx1.8\times10^{-4}\,\mathrm{K^{-1}}$ , tuning ≈70–80 pm/K, loss 1–3 dB/cm, $V_\pi L\sim1$ –3 V·cm) are standard values from Reed and Knights and Chrostowski and Hochberg; the plasma-dispersion relations follow Soref and Bennett (1987), and $\beta_{\mathrm{TPA}}\sim0.5$ –0.9 cm/GW is representative.

Prerequisites. Total Internal Reflection and Optical Guiding · Waveguide Modes and the Wave Equation · Integrated Photonic Waveguides and Circuits · Numerical Aperture and Acceptance Angle. Leads to. Microring Resonators and Optical Cavities · Coupled-Mode Theory and Directional Couplers · Dispersion Management · Single-Mode and Multimode Operation.

Ask the tutor: ask me to walk through how a reverse-biased p–n junction actually produces the $\Delta n_{\mathrm{eff}}$ that switches a silicon Mach–Zehnder, from depletion width to phase.

Why is the index contrast, not silicon's high index alone, what matters?

Confinement is set by the difference between core and cladding index, not the core index by itself. The step $n_{\mathrm{Si}}-n_{\mathrm{SiO_2}}$ fixes the critical angle and numerical aperture, and the term $n_{\mathrm{core}}^2-n_{\mathrm{clad}}^2\approx10$ in the slab eigenvalue drives the evanescent decay constant up. A silicon core clad in another high-index material would confine poorly. The genius of silicon-on-insulator is pairing a very high core index with a low oxide cladding, maximising the step and the confinement.

If silicon can't lase, how do silicon photonic chips get their light?

By importing it. The dominant route is heterogeneous integration: a III–V gain die is wafer-bonded or grown onto the silicon, and its mode is evanescently coupled into a silicon waveguide — light generated in the III–V, routed in silicon. A simpler route keeps the laser off-chip and couples it in through a fibre and a grating or edge coupler. Both accept the same verdict: silicon's indirect gap makes it superb at guiding, filtering, modulating, and detecting light, but hopeless at generating it.

Why does a silicon modulator add absorption when it changes the phase?

Because the only fast index-changing knob silicon offers is plasma dispersion, and free carriers do two things at once. Injecting or depleting electrons and holes shifts the real index (equation 3), but the same carriers also absorb — free-carrier absorption $\Delta\alpha$ , scaling with density as $\Delta n$ does. The two are inseparable, both arising from the Drude response of one carrier gas. This is the price of lacking a Pockels effect: lithium niobate shifts phase losslessly through $\chi^{(2)}$ , whereas silicon must move charge, and moving charge dissipates light.

Why is waveguide loss usually quoted as 1–3 dB/cm rather than something intrinsic?

Because at 1550 nm intrinsic silicon absorption is negligible — the loss is almost entirely scattering from etched-sidewall roughness. A lithographic edge is never atomically smooth, and the scattered power scales with the square of the RMS roughness and steeply with the contrast, since the scattered field is driven by the polarization discontinuity at the boundary. So the same contrast that confines the mode amplifies its sensitivity to roughness. Loss is therefore a fabrication number: smoothing sidewalls by oxidation or etch tuning pushes strip guides toward sub-dB/cm, and lower-contrast silicon-nitride guides are quieter still.

What sets the trade-off between a microring's free spectral range and its Q?

Geometry, pulling both ways. The free spectral range $\mathrm{FSR}=\lambda^2/(n_g L_{\mathrm{rt}})$ grows as you shrink the round-trip length, so a small ring spreads its resonances far apart — good for a filter that must not alias neighbours. But a small ring bends tightly, and tight bends radiate, lowering Q and broadening each resonance. You cannot maximise both range and sharpness; a designer picks a radius giving enough FSR for the channel plan while keeping bending loss low enough for the target Q, then adds a heater to park the resonance.

Why do two 1550 nm photons cause absorption when one cannot?

Pure energy bookkeeping against the bandgap. A single 1550 nm photon carries about 0.80 eV, below silicon's 1.12 eV gap, so it passes through — hence transparency. But two photons together deposit 1.60 eV, above the gap, and can be absorbed in one event. Because it needs two photons at once, the rate scales with intensity squared: invisible at low power, switching on only when the light is intense, as in a tight waveguide or high-Q resonator. Worse, each event leaves a free electron–hole pair that adds its own free-carrier absorption.