GalaxyGen64 — Internship

This repository is the hub that ties together the work I carried out during my end-of-studies internship at the CosmoStat laboratory (CEA Paris-Saclay), from April 2026 to September 2026, supervised by Samuel Farrens and Emma Ayçoberry.

The goal was to build an end-to-end pipeline for generative modelling of galaxy images: assemble a training dataset from real Euclid observations, train a deep generative model on an HPC cluster, and check how faithfully the generated galaxies reproduce the morphology of real ones.

The full write-up is available here: Internship_report_final.pdf.

Project repositories

The work is split across three standalone repositories, one per stage of the pipeline.

1. Dataset construction — Euclid-Q1-postage-stamps

Turns Euclid Q1 VIS calibrated imaging into a machine-learning-ready dataset of 64×64 postage stamps of isolated sources, each paired with its noise map, a bad-pixel mask, and a locally interpolated PSF (optionally a residual PSF kernel). The pipeline has three stages:

• Acquisition — query the Euclid Science Archive; download the calibrated science frames, their background frames, the global VIS PSF model, and the cross-matched source catalogues; slice everything into per-quadrant FITS files.
• Build — for every selected, isolated catalogue source (cuts on flux, point_like_prob, spurious_prob, and bad-pixel fraction), cut a background-subtracted stamp with its noise and mask, interpolate the PSF at the source position, and collect the records into a Hugging Face datasets.Dataset.
• Output — save the dataset locally and/or push it to the Hugging Face Hub.

2. Model training — Train-AE

Trains the two-stage generative model on the 64×64 galaxy postage stamps, run on the Jean Zay supercomputer (IDRIS/CNRS):

1. A convolutional autoencoder that reconstructs galaxy images, convolving its output with the instrument PSF (via jax-galsim) before comparing it to the observed, PSF-convolved image. Trained in a full-PSF variant (with a total-variation term to tame the ill-posed deconvolution) and a partial-PSF variant that avoids the non-physical regularisation.
2. A normalising flow fitted on the frozen autoencoder's latent space, so new galaxies can be sampled by drawing latent codes from the flow and decoding them.

Written in JAX/Equinox, training logged to Weights & Biases. The modelling code builds on prior work by Benjamin Rémy (former CosmoStat PhD student).

3. Result verification — galaxy-morphometrics

A standalone toolkit that checks the generated images by comparing the morphology of real, reconstructed, and flow-sampled galaxies. It reproduces the validation plots of Lanusse et al. (2020) / deep_galaxy_models, computing per-object statistics — HSM adaptive moments (size, ellipticity, rho4), CAS (concentration, asymmetry), Gini/M20, and MID — and rendering their distributions with one curve per dataset (real, reconstruction, flow prior). Train-AE also carries a complementary PQMass check that generated samples match the held-out data distribution in both latent and image space.

Pipeline overview

Euclid Q1 VIS calibrated imaging
        │
        ▼
[Euclid-Q1-postage-stamps]  →  64×64 stamps (+ noise, mask, PSF) on the HF Hub
        │
        ▼
[Train-AE]  →  autoencoder + latent normalising flow, trained on Jean Zay
        │
        ▼
[galaxy-morphometrics]  →  morphology of real vs. reconstructed vs. generated galaxies

Internship details

Laboratory	CosmoStat, CEA Paris-Saclay
Period	April 2026 – September 2026
Supervisors	Samuel Farrens, Emma Ayçoberry
Compute	Jean Zay supercomputer (IDRIS/CNRS)
Data	Euclid Q1 release, VIS instrument