Chapter 27Lasers on Silicon

Silicon cannot lase efficiently, so the light source is brought in: bonded, printed, flip-chipped or grown.

In this chapter you will learn to
  • compare off-chip, flip-chip, heterogeneous, micro-transfer-printed and monolithic lasers;
  • explain how light moves between a III–V gain region and a silicon waveguide through supermodes and tapers;
  • estimate self-heating and the thermal impedance of a laser on silicon-on-insulator;
  • describe open laser-integration projects (the SiEPICfab Shuksan run, UBC ELEC 413);
  • explain how directly modulated microring lasers combine source, modulator and WDM, and what limits them.
Design question

Which integration approach for which volume and performance?

27.1Why the light source is the hard part

Silicon’s indirect gap rules it out as a laser material (Chapter 22), so every silicon photonic transceiver gets its light from a III–V semiconductor (Chapter 23). The source is often the costliest and least reliable part of the link. It is temperature-sensitive (Chapter 24), sensitive to reflections (Chapter 25), and made on small, expensive InP wafers. It is also where most of the link’s power goes, once its wall-plug efficiency is counted (Chapter 33). How the III–V and the silicon are joined decides the cost, the coupling loss, the thermal behaviour and how many lasers fit on a chip [38, 39].

27.2External and flip-chip lasers

  • Off-chip sources. A packaged laser, often in its own module, feeds the silicon chip through a fibre. It is simple, replaceable and keeps the heat away from the photonic chip, and it is the choice of many co-packaged-optics designs (Chapter 17). The cost is fibre-coupling loss at both ends and the space for the module.
  • Flip-chip and butt coupling. A finished laser die is placed face down on the silicon chip, aligned to an edge coupler, either passively against lithographic stops or actively (Chapter 17). Micro-optical benches with lenses and an isolator are a variant. The laser can be tested before assembly, so only good dies are used.
  • Vertical-cavity lasers. A VCSEL (Section 24.8) emits from its surface, so it can be flip-chipped face down over a grating coupler (Chapter 15) after being tested on its own wafer. The 850 nm VCSELs of multimode links are absorbed by silicon, though, so this needs long-wavelength VCSELs, which are harder to make, or silicon-nitride waveguides, which are transparent at 850 nm.

27.3Heterogeneous integration

In heterogeneous integration, unpatterned III–V material, epitaxial layers on an InP wafer or die, is bonded face down onto the processed silicon wafer, its substrate is removed, and the lasers are then defined lithographically, aligned to the silicon waveguides beneath [34]. Bonding uses a thin oxide (molecular bonding) or a polymer such as BCB. Because the lasers are made after bonding, their alignment to the silicon is lithographic, and many lasers, amplifiers, modulators and detectors can be made at once on the full silicon wafer.

The light must move between the III–V gain region above and the silicon waveguide below. The two waveguides form one coupled structure, whose supermodes (Chapter 9) share the light according to how well matched the two guides are. Under the gain section, a narrow silicon waveguide leaves the fundamental supermode in the III–V, where it gets gain. Widening the silicon along a taper raises its effective index past the III–V’s, and the fundamental supermode moves down into the silicon, adiabatically if the taper is long enough. The III–V mesa is then tapered to a tip and ends.

Table 27.1 shows this for a generic stack computed with femwell (figsrc/ch27_hybrid.py): a 500 nm silicon strip under a 3 µm wide InP mesa with a 300 nm multiple-quantum-well region, bonded through 30 nm of oxide.

Table 27.1. Fundamental supermode of a III–V-on-silicon waveguide against the silicon width (femwell, 1550 nm, TE). Confinement is the fraction of the mode’s energy in each region.

Si width neff In the MQW region In the silicon
0.40 µm 3.193 52 % 1 %
0.60 µm 3.198 52 % 4 %
0.75 µm 3.207 46 % 20 %
0.85 µm 3.218 37 % 38 %
1.00 µm 3.236 25 % 60 %
1.50 µm 3.273 14 % 78 %

The transfer happens over a narrow range of widths, around 0.85 µm, where the light is shared equally. A taper that sweeps the silicon width slowly through this range, over tens to a couple of hundred micrometres, carries the light across with little loss. The designer can also choose an intermediate width under the gain section to set the modal gain, trading gain (Γ in the MQW region) against the silicon’s control of the mode, for example for a laser whose feedback comes from a silicon Bragg grating or rings.

Try it. Widen the silicon and watch the computed mode move from the III–V into the silicon; compare the fundamental with its partner.

// try it · III–V on silicon: supermodes as the silicon widens under the gain region (femwell)
geometry: to scale, with the computed mode
taper top view: length squeezed, width stretched
supermode index against silicon width
MQW and silicon share of the selected mode
neff –in the MQW region –in the silicon –field drawn for –

27.4Micro-transfer printing

Micro-transfer printing picks up finished, tested III–V devices (lasers, amplifiers, photodiodes), each released from its source wafer by etching a sacrificial layer, with a soft elastomer stamp, and prints them onto the silicon wafer, many at a time. The devices are then wired up on the target wafer. It combines the known-good-die advantage of flip-chip with the density of bonding, and uses the III–V wafer efficiently. Coupling to the silicon is through evanescent tapers or butt coupling; alignment accuracy of the printing sets the tolerances.

27.5Monolithic growth

The ultimate aim is to grow the laser directly on silicon. The 4–8 % lattice mismatch and thermal mismatch between III–V and silicon create dislocations that kill conventional quantum-well lasers. Quantum-dot active regions tolerate them far better, because carriers are captured by the dots before they can diffuse to a defect (Chapter 23). InAs quantum-dot lasers grown on silicon now run continuously with long lifetimes. The challenge for silicon photonics is coupling: the thick buffer layers that filter dislocations lift the active region micrometres above the silicon waveguides, so growth in recessed pockets, aligned to the waveguides, is an active field [39].

27.6Thermal and reliability

A laser on silicon-on-insulator sits on the buried oxide, a good thermal insulator (Chapter 18). The heat it dissipates, \(P_\mathrm{diss} = IV - P_\mathrm{out}\), raises its temperature by

Self-heating
\[\Delta T = Z_\mathrm{th}\,(IV - P_\mathrm{out}), \qquad I_\mathrm{th}(T) = I_\mathrm{th,0}\,e^{\Delta T/T_0}\]

with a thermal impedance \(Z_\mathrm{th}\) that is larger on SOI than for a laser on its own InP substrate. Heat sinks through openings in the BOX (thermal shunts), and metal heat spreaders lower it.

Worked Example 27.1 · How hot does it get?

Suppose a laser with \(Z_\mathrm{th} = 80\) K/W is driven at 100 mA and 1.5 V and emits 20 mW. It dissipates 130 mW and runs 10.4 K above its heat sink. With \(T_0 = 60\) K its threshold is 19 % higher than if it were not self-heating, its gain peak moves by about 5 nm (and with it a Fabry–Pérot laser’s wavelength; a DFB or ring laser’s moves by about 1 nm), and at higher currents the extra heating can make the output roll over. Halving \(Z_\mathrm{th}\) with a thermal shunt halves all of these.

Reliability follows temperature: III–V lasers age faster when hot, through defect growth in the active region and facet degradation. Bonded and grown lasers must pass the same accelerated-ageing tests as conventional ones, at the highest operating temperature they will see.

27.7Open laser-integration projects

Laser integration has mostly been the domain of companies, but open projects now let students and researchers design lasers on silicon. The SiEPICfab Shuksan process design kit, released under the MIT licence, supports a multi-project run with laser integration [26]. UBC’s ELEC 413 course designs semiconductor lasers on these runs and publishes its designs and measurements [12], a source of real L–I curves and spectra for the analysis of Chapter 32.

27.8Directly modulated microring lasers

Most silicon transceivers modulate a continuous laser externally. A research alternative removes the modulator: make the laser itself a ring on silicon, and modulate its current directly [94]. A heterogeneous microring laser is a III–V ring tens of micrometres across, bonded over a silicon bus waveguide that collects its light. It needs no facets or Bragg mirrors, its small volume gives a low threshold, and its diameter sets its wavelengths. A row of rings of slightly different sizes on one bus is then a WDM transmitter with one directly modulated laser per channel and no separate multiplexer.

The ring’s sensitivity is also its difficulty. A 40 µm ring with \(n_g = 3.7\) has modes 3.69 nm apart at 1310 nm, but making it 0.1 % larger moves every mode by 1.31 nm, so lithography alone cannot place the channels, and each ring needs tuning, by a heater or by a MOS capacitor built into the ring. With an effective-index change of \(2\times10^{-4}\) per kelvin, the modes also drift by 0.07 nm/K as the ring heats with its own drive current, so a thermal shunt (Section 27.6) matters as much as for any laser on silicon. Driving such a laser raises every problem of Chapter 25: its response is fast and ringing on rising edges and slow on falling ones, so its CMOS driver equalises the two edges separately, designed against a compact model of the ring. Quantum-dot gain (Section 27.5) lowers the threshold and steadies it against temperature, and injection locking from an external master laser extends the modulation bandwidth several times.

Answer to the design question

For low volumes and flexibility, an external laser or a flip-chipped known-good die. For high volumes with many lasers per chip, heterogeneous integration, which makes them all at once, lithographically aligned. Micro-transfer printing sits between the two, adding tested devices at high density. Monolithic quantum-dot growth promises the lowest cost at scale, once the coupling problem is solved. Every approach must manage heat on the insulating BOX.

Reference-link choice

Four flip-chipped DFB lasers, tested before assembly and butt-coupled to edge couplers (Chapter 16), each on a thermal shunt through the BOX. A heterogeneous version is the high-volume alternative.

Reading

  • [38] Di Liang and John E. Bowers. “Recent progress in lasers on silicon.” Nature Photonics, 4:511–517, 2010.
  • [34] Tin Komljenovic et al. “Heterogeneous silicon photonic integrated circuits.” Journal of Lightwave Technology, 34(1):20–35, 2016.
  • [39] Zhiping Zhou, Bing Yin and Jurgen Michel. “On-chip light sources for silicon photonics.” Light: Science & Applications, 4:e358, 2015. Open access.
  • [26] SiEPIC program. SiEPICfab Shuksan PDK: the SiEPICfab / Applied Nanotools / Dream Photonics laser-integration MPW run (MIT licence). https://github.com/SiEPIC/SiEPICfab_Shuksan_PDK
  • [12] SiEPIC program, University of British Columbia. UBC ELEC 413 Semiconductor Lasers: design submissions and measurement data (MIT licence). https://github.com/SiEPIC/UBC-ELEC413-2025Fall
  • [94] Sam Palermo. “Lecture 12: Laser Sources.” ECEN 721: Optical Interconnects Circuits and Systems, Texas A&M University, Spring 2026. https://people.engr.tamu.edu/spalermo/ecen689_oi/lecture12_ee721_laser_sources.pdf
Key takeaways
  • Lasers reach silicon off-chip, flip-chipped, bonded, transfer-printed or grown; each trades cost, density, coupling and heat.
  • In heterogeneous lasers the fundamental supermode moves from the III–V to the silicon as the silicon widens: in the example, from 52 % in the MQW region and 1 % in silicon at 0.4 µm to 14 % and 78 % at 1.5 µm, crossing near 0.85 µm.
  • Adiabatic tapers sweep through the crossing to transfer the light with little loss.
  • Self-heating, ΔT = Zth(IV − Pout), is worse on the BOX; thermal shunts help.
  • Open PDKs and course data (SiEPICfab Shuksan, UBC ELEC 413) make laser integration accessible.
  • Surface-emitting VCSELs can be flip-chipped over grating couplers, but only long-wavelength ones suit silicon waveguides.
  • Directly modulated microring lasers give one laser per WDM channel without modulators (modes 3.69 nm apart for a 40 µm ring), at the price of tuning (1.31 nm per 0.1 % of diameter), self-heating and edge-dependent dynamics.