docs: draft end-to-end CASA → PyAutoLens data-prep script

[Jammy2211]

Overview

The existing scripts/interferometer/casa_reduction.py is an incomplete stub with most
code commented out. Interferometer users on Slack repeatedly need guidance on turning
CASA-reduced ALMA .ms data into PyAutoLens-compatible visibilities, uv-wavelengths
and sigma arrays. This task produces a first-draft end-to-end walkthrough script that
can be iterated on with real users in Slack.

Plan

• Research CASA reduction workflows and ALMA .ms structure
• Read existing interferometer workspace scripts for convention
• Rewrite casa_reduction.py as a coherent narrative covering split → channel avg →
  (optional uvcontsub) → statwt → export → build al.Interferometer
• Borrow working helpers from z_staging/interferometer_data_prep/
• Aim for "good enough to share on Slack", not final

Detailed implementation plan

Affected Repositories

• autolens_workspace (primary)

Work Classification

Workspace

Branch Survey

Repository	Current Branch	Dirty?
autolens_workspace	main	dirty (unrelated dataset/config changes from prior session)

Suggested branch: feature/interferometer-data-prep
Worktree root: ~/Code/PyAutoLabs-wt/interferometer-data-prep/

Implementation Steps

1. Read scripts/interferometer/start_here.py, data_preparation.py, simulator.py
   to match tone, imports (import autolens as al), and docstring style.
2. Rewrite scripts/interferometer/casa_reduction.py:
   • Intro: .ms layout (2, n_spw, n_c, n_v, 2) vs al.Interferometer shapes
   • Section: CASA split per field/spw
   • Section: channel averaging via split(width=...)
   • Section: continuum subtraction (uvcontsub) — optional
   • Section: statwt for realistic sigmas
   • Section: Python helpers for DATA/UVW/CHAN_FREQ/SIGMA export to FITS
     (pulled from z_staging/interferometer_data_prep/main_uvcontsub_SPT0538_CO9-8.py)
   • Section: Concatenate spws, average polarisations, build al.Interferometer
   • Section: Slack pointer for edge cases
3. Regenerate notebook via /generate_and_merge after user review.

Key Files

• scripts/interferometer/casa_reduction.py — full rewrite
• notebooks/interferometer/casa_reduction.ipynb — regenerated later

Original Prompt

Click to expand starting prompt

In the file @autolens_workspace/scripts/casa_reduction.py, I attempt to give a run
through of how one can prepare data via CASA.

In the file @z_staging/interferometer_data_prep/main_uvcontsub_SPT0538_CO9-8.py, I
have example reduction scripts from a an autolens user.

Can you do a bit of deep research on CASA and interferometer data reduction in
general, read a few interferometer workspace scripts for context and attempt to
produce some sort of useable and help data reduciton script for interferometer
users? It doesnt need to be perfect, I will back and forth it with interferometer
people on the autolens SLACK.

[Jammy2211]

Shipped — 2026-04-15

PRs

• docs: draft end-to-end CASA → PyAutoLens data-prep script #69

Summary

Rewrote scripts/interferometer/casa_reduction.py from an incomplete stub into a coherent end-to-end walkthrough. Covers the 5-step CASA pipeline (split → channel avg → uvcontsub → statwt → FITS export), Python helpers borrowed from a working ALMA SPT-0418 user script, polarisation-averaging + SPW concatenation ordering, and the canonical al.Interferometer.from_fits load path with a small fast real-space mask for iteration.

Session Notes

• On-disk FITS shape correction: al.Interferometer.from_fits expects (n_vis, 2) real-valued FITS (col 0 = real, col 1 = imag) for both data and noise-map — NOT (n_vis,) complex. First-draft wording said complex; corrected after reading Visibilities.from_fits in PyAutoArray.
• Pol-average-first ordering: polarisations must be averaged before channels are flattened into n_vis, otherwise pols end up interleaved with channels and the noise-map stops corresponding to its visibility. Spelled out explicitly in the script.
• get_sigma robustness: CASA's np.squeeze drops the polarisation axis when n_pol == 1; added a guard to re-introduce it so single-pol reductions don't crash.
• Real-space mask sized for speed: example mask is (64, 64) @ 0.05"/px, radius 1.5" so a dirty-image sanity check finishes in seconds; production numbers (256x256 @ 0.02") are called out in a follow-up paragraph.
• Stale notebook: notebooks/interferometer/casa_reduction.ipynb is out of date. Deferred to /generate_and_merge since in-worktree notebook regen picked up 48 cross-repo changes.
• Future work: Slack feedback will drive whether we extend this to mosaicking, heterogeneous SPWs, and joint-polarisation modeling.