The Laser That Cannot Fall Out Of Lock

An integrated laser with a sub-10-hertz linewidth is not new. An integrated laser with a sub-10-hertz linewidth that stays there no matter what current you put through it, with no feedback loop, no dither, no lock acquisition and no engineer watching an oscilloscope — that is new, and it is what a group at EPFL reported in Nature Photonics on 28 September. The device holds an intrinsic linewidth below ten hertz at all drive currents, with no active control anywhere in the apparatus.

The distance between those two sentences is the distance between a result and a product.

Start with the problem. A semiconductor laser diode is small, efficient, electrically pumped and manufacturable by the million, and its frequency is appalling. Spontaneous emission events inside the cavity kick the phase of the output field at random, and because the diode's cavity is short and lossy, each kick is large relative to the stored energy. The result is a linewidth measured in megahertz — fine for intensity-modulated datacom, useless for anything that treats the optical phase as a measurement. Optical clocks, fibre sensing, coherent communication and frequency-modulated lidar all need the phase to behave.

Self-injection locking is the trick that fixes this without adding a second laser. Couple the diode to an external optical resonator with a very high quality factor. A tiny fraction of the emitted light is returned to the diode by backscattering inside that resonator — in a silicon nitride microring or spiral, by Rayleigh scattering off the sub-nanometre roughness of the waveguide sidewall, which couples the forward travelling mode into the backward one. That returned field is spectrally narrow, because the resonator is narrow, and when it re-enters the diode it forces the diode to oscillate at the resonator's frequency rather than its own. The diode inherits the resonator's spectral purity. In the EPFL device the spiral resonator has an intrinsic quality factor of 11.3 million, and the locked laser's frequency noise falls more than 5,000-fold below the free-running diode, with the intrinsic linewidth reduced by a factor of at least 10⁴.

The reason is phase. The feedback field does not simply return to the diode; it returns having accumulated a round-trip phase, set by the optical path length between the diode facet and the resonator. Whether that returning field reinforces the diode's oscillation or fights it depends on that phase. Locking happens over a limited window of feedback phase — roughly, where the returned field pulls the diode frequency towards the resonance rather than pushing it away. Outside the window there is no stable locked state, and the laser falls back to its free-running noise or, worse, hops chaotically. In this device the locking range is widest at the most preferential feedback phase, where it spans 6.7 gigahertz.

A gap of a few micrometres between a diode facet and a chip edge is a phase that moves with everything. Drive current changes the diode's junction temperature, which changes its refractive index and its emission frequency; the chip warms, the gap changes by a fraction of a wavelength, the feedback phase slides. Tune the current to tune the laser and the feedback phase walks straight out of the locking window. In practice this has been handled by putting a heater in the feedback path and servoing it, or by accepting that the laser only works at a handful of blessed currents found by hand. Either way the compact, manufacturable, low-noise laser arrives with a control system attached, and the control system is larger and more expensive than the laser.

The EPFL answer does not fight the phase drift. It arranges for the drift to always land somewhere good.

The key is that the laser does not see one resonance. It sees a comb of them, spaced by the resonator's free spectral range, and the EPFL resonators are long spirals with unusually small free spectral ranges — 1.44 gigahertz in one device, 1.5 gigahertz in another. That spacing is close enough that as the diode's own frequency sweeps with current, it crosses resonance after resonance. The question is what the feedback phase is doing at each successive crossing.

Get the sign and size right, and something satisfying happens. As the current tunes the laser and the feedback phase slides out of the locking window at one resonance, the next resonance in the comb arrives with its phase already placed inside the window. The laser hands itself from one resonance to the next without ever passing through an unlocked state. There is no acquisition, because it never loses lock; there is nothing to servo, because the geometry has already done the servo's job. The authors call it endless self-injection locking, and they put a number on the condition: it holds for feedback-phase dispersion between 0.15π and 1.35π. Negative dispersion does the opposite — the phase runs away from the window, and the laser unlocks between resonances.

A fair analogy is a spiral staircase with a handrail that is only within reach for part of each turn. Conventional self-injection locking climbs one turn and loses the rail. The EPFL design offsets the rail slightly on each turn so that it returns to the hand just as the previous grip is lost. The climber never has to stop and feel for it.

The device is an off-the-shelf distributed-feedback diode emitting at 1,548 nanometres — the telecom C-band, where every component is cheap — butt-coupled to a silicon nitride photonic integrated circuit carrying the spiral resonator. The resonator's intrinsic loss rate κ₀/2π is 17.1 megahertz, against bus and drop coupling rates of 17.6 and 18.2 megahertz. The sum of the two external rates is roughly twice the intrinsic loss, so the resonator runs well overcoupled, and the loaded quality factor sits a long way below the intrinsic 11.3 million — which is the price paid for getting enough light back out to the diode.

Three resonator architectures were fabricated: one with no drop port, one with a symmetric drop port and an extended bus waveguide, and one with an asymmetric drop port. The device with feedback-phase dispersion of 0.16π behaves as the theory says. Spectrograms taken as the current is swept show the laser locked through feedback-phase period after feedback-phase period, with no unlocked bands. A device with −0.57π shows exactly the unlocking regions the model predicts. That pair is the experiment, and its limits should be stated plainly: the two devices also differ in free spectral range, so dispersion is not the only variable, and the Methods note that the original DFB diode was broken by operator error partway through and replaced by a chip-on-submount part from Sivers for some of the later measurements.

The frequency noise was measured with a delayed self-heterodyne setup — five kilometres of fibre delay and a 40-megahertz acousto-optic modulator — as the injection current was scanned from 154 to 300 milliamps in two-milliamp steps. The intrinsic linewidth, which is the Lorentzian floor extracted from the white-noise part of the frequency-noise spectrum and the number that matters for coherent detection, stays below ten hertz throughout, with nothing stabilising the feedback phase.

The honest caveat is the output power: about 0.7 milliwatts, calibrated as an average over a 100-to-300-milliamp sweep. That is enough for a local oscillator, a clock interrogation beam or a seed, and it is not enough for a transmitter. Anything that needs to put light into the air or down a fibre will need a semiconductor optical amplifier after this laser, and an amplifier adds amplified spontaneous emission, cost, power and another thing to integrate. A 0.7-milliwatt source is a component, not a system, and the paper does not claim otherwise.

The second result is the one that will interest anyone building frequency-modulated continuous-wave lidar. Piezoelectric actuators integrated monolithically on the photonic chip strain the resonator, shifting its resonance and dragging the locked laser with it. Driven with a triangular waveform at 10 kilohertz and 50 volts peak to peak, the laser produces mode-hop-free frequency chirps of 1.5 gigahertz — more than one free spectral range — and the paper reports the mean square deviation from an ideal ramp as a measure of how linear the sweep is.

Strip the claim back to what the data supports and it is still substantial. The specification that previously read "sub-10-hertz linewidth at selected operating points, with active feedback-phase stabilisation" now reads "sub-10-hertz linewidth, any current, open loop". For a system integrator, the second is not an improvement on the first; it is a different product. The heater driver, the photodiode, the error-signal electronics, the acquisition routine and the engineer who writes it all come off the bill of materials, and with them goes much of the reason these lasers have not appeared in commercial instruments.

What has not changed is everything downstream. The diode and the photonic chip are still butt-coupled, and butt-coupling is still where hybrid integration gets expensive — submicron alignment, active placement, and an epoxy or solder joint that has to hold it through thermal cycling. The output power still needs amplification for most uses. The piezoelectric actuators need 50 volts, which is a drive-electronics problem rather than a physics problem but is a problem nonetheless. And three devices on a laboratory bench are not a yield statistic; the claim that feedback-phase dispersion can be set lithographically is plausible precisely because it is geometric, but it has not yet been demonstrated across a wafer.

Still, the thing that was fragile was fragile for a specific reason, and that reason has been removed by changing a geometry rather than by adding a control loop. That is the kind of fix that tends to survive the trip from paper to product.