Three ways to catch a beam

A high-rate satellite-to-ground laser link must put turbulent light into a single-mode fiber. The established receiver does this with a deformable mirror and a wavefront sensor in front of the fiber.

Two other designs are now being built. In the first, a mode sorter splits the incoming beam into many Gaussian channels, and a photonic chip recombines them into one fiber. Cailabs has reported a 45-channel version for ground-station telescopes larger than 40 cm. In the second, an array of small telescopes acts as one aperture. When each aperture is smaller than the turbulence coherence length, setting the relative phase between apertures acts as a form of digital adaptive optics that can scale to many apertures.

The optics differ. The control problem converges.

What the loop now has to do

  • Hold more settings at once. A 45-channel recombiner has tens of phases to keep aligned. An array adds pointing and phase for every aperture.
  • Work from less direct measurements. These receivers often measure power, not the wavefront. The controller must infer the correction from indirect, noisy signals.
  • Act faster. A low-Earth-orbit satellite crosses the sky in minutes. Its motion sweeps turbulence across the aperture, so corrections must arrive at kilohertz rates.
  • Handle changing conditions within one pass. Phase errors dominate on some paths and intensity fluctuations on others. Cailabs saw both regimes on two test links.

The field already points at the loop

Recent work names control, not only optics, as a limit.

Cheaper hardware, smarter loops

These receivers replace one expensive corrector with many fast, inexpensive channels. That trade moves performance from the precision of the hardware to the quality of the control.

A hand-tuned loop per channel does not scale to dozens of coupled channels, indirect measurements, and conditions that change within a pass. A controller that carries a model of the receiver can.

What we are building

Noether Labs builds learned controllers for these receivers, trained on differentiable models of the optics. The same model runs beside the controller in operation. It predicts what the sensors should see. When the instrument departs from the prediction, control returns to the existing classical loop.

We hold this work to the classical baseline. In our first simulated comparison, a learned controller did not outperform both classical baselines in any of twelve conditions (results). That is the bar a learned loop must clear on real telemetry, and it is why the fallback exists.

Our first step with ground-station teams is offline and read-only: build a model from logged telemetry, compare controllers against the existing loop, then run in shadow mode, where the controller records the command it would send and never sends it.

If you operate a ground station or build optical receivers, we would like to hear from you.

Salma Shaik
Founder & CEO, Noether Labs

Optical-links research ↗