AO Wavefront Lab

Biswajit Jana / independent learning project / 2026

Learning adaptive optics.
Following the evidence.

How do wavefront measurements become a trustworthy scientific result? I explored this through released ESO CIAO telemetry, a reproducible sampling audit, and a separate wavefront-control model.

Read the project report ↗Explore the telemetry →

One released observation · approximately 30 seconds · six modelled aberration modes

15,000source frames
300browser samples
79,200telemetry values checked
499.962 Hzreleased loop rate
01 / Starting point

Why I wanted to study this

Adaptive optics connects the subjects I want to understand together: astronomical instrumentation, wavefront sensing, and feedback control.

My aim was to build intuition from real telemetry while learning how an optical model and controller behave. Working through a public AOT dataset gave me a practical route from file formats and sensor streams to numerical validation and the limits of a claim.

The project’s contribution is the software and audit built around an existing release. The observations, instrument, and AOT standard belong to their original teams.

02 / What I built

A traceable path from FITS to interpretation.

The telemetry pipeline and teaching model answer different questions. The model does not ingest CIAO measurements as reconstructed pupil phase.

  1. 01

    Retain the source

    Official CIAO1 AOT FITS file, published MD5, local SHA-256, metadata, and attribution.

    Provenance record ↗
  2. 02

    Prepare a browser product

    Every 50th frame; timestamps and indices retained; six-decimal JSON values for four streams.

    Conversion code ↗
  3. 03

    Test the reduction

    Exact source parity, mean and percentile distortion, and full-rate power above the reduced Nyquist limit.

    Machine-readable audit ↗
  4. 04

    Inspect the result

    Worker-prepared telemetry maps, sampled trends, frame inspection, and explicit interpretation limits.

    Open the observation ↗
Separate learning branch

Noll modes → delayed feedback → residual phase → FFT intensity. An opt-in teaching model supports controlled experiments. It is not a reconstruction of the CIAO controller or telescope PSF.

Model guide →
03 / Measured outcome

What survived the reduction?

All 79,200 retained telemetry values, 300 timestamps, and 300 source indices pass the declared parity tolerance. The three scalar diagnostics also pass the study’s mean and percentile thresholds.

Each diagnostic exceeds the 5% temporal acceptance limit. These are study-specific thresholds, not universal AO standards. A gate failure here limits the browser product; it does not indicate a fault in the telescope.

Read the quantitative interpretation →
Above reduced Nyquist power: gradient RMS 20.89%, command RMS 9.19%, mean flux 64.46%; each exceeds a 5% study limit.
Hann-windowed, mean-centred full-rate diagnostic power above the reduced 4.99962 Hz Nyquist. The full-rate FITS, rather than the 300-frame product, underlies this calculation.
04 / Learning outcomes

What the work helped me understand

Data engineering

A plot needs a chain of custody.

FITS extension shapes, source-frame indices, timestamps, hashing, and rounding tolerances turn a visual demo into an inspectable data product.

Scientific reasoning

Different claims need different tests.

A sample can reproduce a distribution and still lose temporal content. Numerical integrity, representativeness, and physical interpretation require separate evidence.

Optics and control

A model’s assumptions are part of its result.

Noll normalization, sampled-pupil FFTs, delay, noise, and the exact feedback recurrence determine what the model can teach. Hardware calibration is still required for telescope inference.

05 / Next research question

A concrete next research question

How should a compact telemetry product be designed for a specified scientific use?

My proposed next experiment compares the current point sample with an explicitly filtered reduction and a full-rate reference. I would define the target bandwidth, filter response, edge handling, and acceptance tests before running the comparison. This is proposed work.

Start here / 2 minutesProject at a glance ↗Purpose, pipeline, findings, learning outcomes and next steps. Technical review / 10 minutesProject report ↗Methods, equations, evidence, learning record, and a reproducibility route. Inspect and reproduceSource repository ↗Code, original retained data, tests, receipts, and generated evidence.