Keeping an On-Chip Laser Stabilizer-Isolator Locked Without the Controller Adding Noise

A recent silicon nitride chip does two jobs that normally need two bulky parts. It makes a cheap semiconductor laser quieter, and it blocks light reflected back into that laser. It works because the waveguide layout is deliberately lopsided. Light circulating one way around a tiny ring resonator shifts the resonance of light circulating the other way, so the ring has a different resonant frequency in each direction. The catch is that the device only works at a sweet spot of laser power and laser-to-ring detuning. Temperature drift and power changes push it away from that point, and nothing on the chip pushes back. This note proposes adding a feedback loop that holds the device at its sweet spot. The loop is designed around one fact: its own small, rapid corrections must not reappear as new laser phase noise or as flickering isolation.

The question we want to answer is concrete and measurable. Robust digital controllers correct the system in small rapid jerks, a behavior called chattering. Does that chattering turn into laser phase noise? We expect the answer to depend mostly on where the controller acts, relative to the passive optical lock, and much less on which control law it uses. A loop that drives the laser's own current should have its jerks compressed by the optical lock. A loop that drives the ring's heater should pass them straight through to the output. If this holds, it gives a simple design rule. It would also give the first actively held version of the device, which reports which sensor signal is worth adding and which controller is worth its complexity.

Background and the gap

The device is the "Unified Laser Stabilizer" of Unified laser stabilization and isolation on a silicon chip. It routes part of the forward laser light back through a high-quality ring in the opposite direction. The material's Kerr effect (a refractive index that depends on light intensity) then acts unequally on the two directions. The forward beam shifts the backward resonance twice as strongly as it shifts its own, which splits the two resonances. The forward laser locks itself to the forward resonance, a process called self-injection locking, in which light fed back from the ring pulls the laser onto the ring frequency. External reflections arrive off-resonance and are dumped. The reported results are a linewidth reduced from 51 kHz to 89 Hz, 25 to 35 dB of frequency noise reduction, and 14 to 15 dB of isolation at about 45 mW on chip. The device is entirely passive. Because the resonance split scales with circulating power, any power or detuning change also changes the isolation depth. This paper motivates the proposal and supplies the platform, but it contains no feedback loop at all.

Active loops elsewhere in the file are all linear. Realization of Hollow-Core Photonic-Crystal Fiber Optic Gyro Based on Low-Noise Multi-Frequency Lasers with Intermediate-Frequency Difference locks a beat note with a field-programmable gate array (FPGA) running a proportional-integral loop that drives an acousto-optic modulator. It reports about 40 dB of noise reduction and a beat linewidth below 1 Hz. A coherent nanomechanical oscillator driven by single-electron tunnelling uses feedback to compress a nanotube oscillator's linewidth from about 3 kHz to below 2 Hz. Both show that well-tuned linear feedback is a serious tool for phase noise. They therefore set the bar our proposal must clear rather than leave a gap for it.

On the control side, Precise Motion Control of Wafer Stages via Adaptive Neural Network and Fractional-Order Super-Twisting Algorithm and Fractional-Order Variable-Gain Super-Twisting Control With Application to Wafer Stages of Photolithography Systems supply the main method. Sliding-mode control drives the error onto a target surface by switching hard between correction directions. The super-twisting variant hides that switch inside an integral, so the actuator sees a continuous command. Feedback-based Digital Higher-order Terminal Sliding Mode for 6-DOF Industrial Manipulators supplies a constraint: digital sampling makes some residual limit cycle unavoidable even with higher-order sliding modes. Sliding-Mode Control for PMLSM Position Control—A Review supplies another: smoothing the switch with a boundary layer gives up exact disturbance rejection. Asymmetric Constrained Control of a Cervical Orthotic Device Based on Barrier Sliding Modes shows how to build unequal bounds into a sliding surface. We reuse that idea because the device's safe operating window is itself asymmetric in detuning, as described in the next section.

None of these papers places a robust, chattering-prone controller around a nonreciprocal optical lock. None asks how the controller's residual oscillation propagates into phase noise and isolation.

Core Claim

Inside its locking range, a self-injection lock acts as a passive smoothing layer. A change $\delta\nu_{\text{free}}$ in the free-running laser frequency, for example from a step in drive current, appears at the output reduced to

$$\delta\nu_{\text{out}} \approx \delta\nu_{\text{free}}/K,$$

where $K$ is the lock's stabilization factor. The reduction comes from the standard locking picture, in which optical feedback pins the laser to the ring and attenuates frequency excursions by the same factor that suppresses its noise. By contrast, a shift $\delta\nu_{\text{ring}}$ of the ring resonance, for example from the heater, drags the locked laser with it almost one-for-one.

The reported 25 to 35 dB reduction in frequency noise power implies $K$ of roughly 18 to 56 in frequency amplitude. This is our inference from the reported range, not a number the paper states.

The claim has three parts. First, a robust controller that acts on laser current can tolerate chattering amplitudes about $K$ times larger than one acting on the ring heater before it raises the phase-noise floor. Second, the controller should regulate a direct measurement of isolation as well as frequency error. That measurement is light leaking out of the backward dump port, which tracks the Kerr resonance split. Regulating it lets the loop hold isolation steady under power changes, which a frequency-only loop cannot do. Third, the safe window is asymmetric. On one side of the operating point the device loses the lock and the laser jumps to a different mode, called a mode hop. On the other side it only loses isolation. An asymmetric barrier sliding surface can therefore use low gain where the margin is wide and act firmly only near the mode-hop edge.

Research Program

Simulation model. We will first build a numerical model of the full device. It combines semiconductor laser rate equations with optical feedback, coupled-mode equations for the forward and backward ring fields including both direction-dependent Kerr effects, a first-order thermal model of the ring, and a sampled controller with realistic latency. The model will be calibrated to reproduce the reported passive results: the 51 kHz to 89 Hz narrowing, 25 to 35 dB of noise reduction, and 14 to 15 dB of isolation.

Checking the compression law. Within the model we map the locking range and the isolation-versus-detuning curve. We then inject controlled square-wave dithers on each actuator to measure the transfer from actuator chatter to output frequency noise. This tests the $1/K$ compression law directly.

Controllers. We implement four controllers on two actuators, laser current and ring heater:

  • a proportional-integral loop with anti-windup;
  • a boundary-layer sliding-mode controller;
  • a super-twisting controller;
  • an asymmetric barrier super-twisting controller.

Each is run with two error signals: frequency error alone, and frequency error plus the backward-port isolation readout. The frequency error comes from a Pound-Drever-Hall-style demodulation or from the ring's transmission dip. The controllers will run on an FPGA at sampling rates between 100 kHz and a few MHz, where the file's concerns about computing cost at high sampling rates apply. Fractional-order terms and neural-network disturbance estimators are excluded unless the simpler controllers fail.

Hardware. We then move to a fabricated stabilizer chip of the published design, with the same disturbances replayed.

Baselines

Passive device, no loop. The naive reference point is the passive stabilizer with no feedback. We operate it at the same power and temperature under the same disturbance scripts. Its reported static numbers are the starting point: 89 Hz instantaneous linewidth, 25 to 35 dB frequency noise reduction, and 14 to 15 dB isolation, all from Unified laser stabilization and isolation on a silicon chip. No reported result for its behavior under drift or power steps was found on Lacuna, so we will measure that ourselves. Beating this baseline means holding lock and keeping isolation within 1 dB of its peak through disturbances that make the passive device mode-hop.

Tuned linear loop. The strongest existing method for this kind of job is a digital proportional-integral loop with anti-windup on the same actuators and signals. This is the architecture of Realization of Hollow-Core Photonic-Crystal Fiber Optic Gyro Based on Low-Noise Multi-Frequency Lasers with Intermediate-Frequency Difference, which reports about 40 dB of beat-note frequency noise reduction and a beat linewidth below 1 Hz on its own platform. No reported result was found on Lacuna for any active loop on a Kerr-nonreciprocal stabilizer. The tuned linear loop is the threat we take most seriously because it produces no chattering by construction.

For fairness, every controller gets the same FPGA, sampling rate, actuator bandwidth, and an equal automated tuning budget. That budget is a fixed number of Bayesian-optimization trials on the same cost: integrated frequency noise plus isolation variance plus a penalty for losing lock. Beating the linear loop means a higher lock-hold rate or smaller isolation excursions under fast steps, with integrated phase noise no worse than 1 dB above it.

Smoothed sliding-mode controller. This baseline replaces the hard switch with a saturation function, as in Parallel-Connected Inverter Control Method in Material Micromachining. It tests whether simple smoothing already captures any robustness gain.

Ablations. Two ablations isolate the proposal's key ingredients. The first runs the identical super-twisting controller on the ring heater instead of the laser current, which isolates the claimed $1/K$ attenuation of chatter. The second removes the backward-port isolation readout, which isolates the value of sensing the asymmetry directly. A related precedent from the control literature is the wafer-stage result, where super-twisting plus a learned disturbance estimator reached 4.95 µm RMS error against 7.43 µm for PID and 9.12 µm for standard sliding mode. That comparison comes from Precise Motion Control of Wafer Stages via Adaptive Neural Network and Fractional-Order Super-Twisting Algorithm and was measured on a mechanical stage. It motivates trying super-twisting here but does not transfer to optics as evidence.

Why not just…?

Why not just use a proportional-integral loop and stop there? It may be enough, and the experiments are built to find out. Linear loops reach sub-hertz beat linewidths in Realization of Hollow-Core Photonic-Crystal Fiber Optic Gyro Based on Low-Noise Multi-Frequency Lasers with Intermediate-Frequency Difference. Their known weakness is fast, large disturbances near a nonlinear edge. Phase-locked Loop-based Proportional Integral Control for Spiral Scanning in an Atomic Force Microscope shows a linear phase-locked loop losing fidelity above 60 Hz. Our device has exactly such an edge: the mode-hop boundary, together with Kerr bistability. If the tuned linear loop matches the robust controllers on lock-hold rate under the power-step and reflection-burst scripts, we will report that the passive lock removes the need for robust control. The actuator-placement rule and the isolation readout would then remain as the contribution, applied to the linear loop.

Why not just stabilize the chip temperature and laser power separately? A thermoelectric cooler and a power servo remove slow drift. They cannot respond to back-reflection bursts or fast changes in circulating power, which move the Kerr split directly, as noted in Unified laser stabilization and isolation on a silicon chip. We include this configuration as a baseline. If it holds isolation through the fast disturbance scripts, the second ablation will show that the isolation readout adds nothing.

Why not just smooth the sliding-mode switch with a boundary layer? Smoothing trades exact disturbance rejection for a steady offset, according to Sliding-Mode Control for PMLSM Position Control—A Review. Our core claim suggests a particular way this alternative could compete. If the actuator drives laser current, the optical lock already supplies a physical smoothing layer, so a narrow electronic boundary layer might suffice. The smoothed controller is therefore run on both actuators, and we report whether its advantage over hard switching disappears when it drives laser current.

Proposed Experiments

Test 1: chatter transfer. This test checks the compression law. On the laser-current actuator, the ratio of output frequency deviation to injected square-wave dither should track $1/K$ across the locking range and collapse near its edges. On the heater, the ratio should stay near one. A transfer ratio similar for both actuators would falsify the core claim.

Test 2: disturbance scripts. All controllers face the same disturbances, both in simulation and on the chip:

  • temperature ramps;
  • 10 to 20 percent laser-power steps;
  • controlled back-reflection bursts from a variable reflector.

Metrics are taken over 100 trials per script:

  • frequency noise spectrum and integrated phase noise;
  • height of spurs at the controller's limit-cycle frequency;
  • isolation over time, measured as backward-to-forward transmission;
  • lock-hold probability.

Test 3: asymmetric surface. This test checks whether the barrier sliding surface lowers average control effort and spur height compared with the symmetric super-twisting controller, at equal lock-hold rate.

Expected outcome. The proposal succeeds if the robust controller on laser current holds lock through disturbances that defeat the passive device. It must also keep isolation within 1 dB and stay within 1 dB of the tuned linear loop in integrated phase noise, with heater-actuated chatter visibly raising the noise floor. If instead the linear loop matches everything, the result is a clean negative finding. It would tell builders of nonreciprocal on-chip lasers to place their actuators on the laser and sense isolation directly, and that robust control is not needed.