The control problem

Atmospheric turbulence distorts the wavefront reaching a telescope. A wavefront sensor measures that distortion; a controller calculates a correction; a deformable mirror applies it. During sensing and computation, the atmosphere continues to change.

We study the trade-off between optical-model fidelity, estimation error, and latency. A more detailed model is useful only when the information it adds changes the control decision enough to justify its cost.

What the system models

Noether AO combines frozen-flow turbulence, wave propagation, deformable mirrors, and pyramid wavefront sensing in JAX. Classical, predictive, and learned controllers encounter shared disturbances, delays, sensor noise, and injected faults.

Independent optical references use HCIPy and SPECULA. Comparisons declare their pupil geometry, normalization, sampling, modulation, and calibration conventions rather than assuming that two simulator defaults describe the same instrument.

Retained results

StudyResult and scope
Prediction Boundary ↗600 held-out simulated trials. Prediction beat the tuned integrator in five of twelve conditions; seven comparisons were uncertain. Complete paired results and a reproduction notebook.
Ideal pyramid comparison ↗36 image cases, 30 derivative cases, and 24 deliberate error controls passed the declared gates. Agreement covers a matched finite-grid ideal model.
Sensor calibration ↗Matched Noether and HCIPy response matrices agree to a maximum relative L2 difference of 5.99 × 10⁻⁹. Native SPECULA calibrations are not directly interchangeable.
Propagation timing ↗CPU component p99: 0.0593 ms at N = 32 and one wavelength; 6.3801 ms at N = 256 and five wavelengths. These are component timings, not complete control-loop latency.

These checks make the model inspectable. They do not establish controller superiority. In the retained baseline study, learned weights beat both the integrator and forward controller in 0 of 12 conditions. Scheduled fallback measurements also remain inconclusive as evidence of safe recovery.

The next result we are working toward

A measured fidelity–latency frontier for a specified pyramid-sensor configuration: actual inference time versus photon flux and held-out closed-loop wavefront error, with tuned classical baselines and paired uncertainty.

The instrument-validation program adds sensor-to-actuator timing, calibration drift, actuator limits, and the cost of switching controllers. The objective is to identify which corrections remain useful under the constraints of a real loop.

Code, methods, and context

The linked papers and external libraries are credited research context. Their authors and institutions are not implied to endorse Noether’s results.