Differentiable adaptive-optics simulation and controller comparisons in JAX.
Noether AO models wavefront propagation, frozen-flow turbulence, deformable mirrors and pyramid wavefront sensing. It compares classical, predictive and learned controllers under shared delays, sensor noise and injected faults. Independent optical references use HCIPy and SPECULA.
Requires Python 3.11 and uv.
git clone https://github.com/noetherlabs/noether-ao.git
cd noether-ao
uv sync --frozen --python 3.11
.venv/bin/noether-optics validate --config configs/reproduction_smoke.yaml --out runs/validationThis reduced example checks the experiment pipeline. Full experiment commands and optional dependencies are in the reproduction guide.
| Experiment | Methods | Results |
|---|---|---|
| Pyramid optics and derivatives | Optical model | HCIPy comparison |
| Wavefront-sensor calibration | Calibration protocol | HCIPy/SPECULA comparison |
| Closed-loop response identification | Estimators | Identification, filtered instruments |
| Scheduled controller handover | Measurement design | Paired trials |
| Optical-gain detection | Frozen protocol | 240-run validation: evaluation withheld |
| Control, noise and faults | Physical criteria | Baseline comparisons |
| Claim | Measured verdict |
|---|---|
| C1 | Conditional support. N=32, one wavelength, float32: p99 0.0592937 ms; N=256, five wavelengths: p99 6.38013 ms. CPU component only. |
| C2 | No tested flux gave useful equivalence across all fidelity levels. No supported low-flux threshold was found; lower fidelity often performed better under this fixed estimator. |
| C3 | Learned weights beat both integrator and forward in 0/12 conditions and lose to at least one in 5/12; remaining comparisons include uncertainty. |
| C4 | Partial detection/recovery, no general safety result. Optical-gain drift missed; vibration fallback worsened WFE; clean false alarms occurred. |
These are simulation and CPU component measurements. They do not establish instrument performance, full sensor-to-actuator latency or safe fallback. The results report gives methods, uncertainty and limitations.
The source is publicly available. No open-source license has been granted for project-owned code; third-party notices retain their stated terms.