25–27 nov. 2026
IJCLab
Fuseau horaire Europe/Paris

Deblending real galaxies from EDP2 using DebVAder

27 nov. 2026, 11:40
10m
IJCLab

IJCLab

Auditorium Joliot-Curie, 100 A, 91400 Orsay, France

Orateur

Lia Lammert (APC / CNRS)

Description

In this work, we focus on DebVAder (Arcelin et al, 2021), a python package built on a variational autoencoder ML architecture to deblend and reconstruct isolated galaxies. Its feasibility has been explored but is limited to simulated albeit realistic data.

As a first test on real data, we apply the simulation-trained DebVAder architecture on the LSST EDP2 dataset from the ECDFS field. We assess the ability of the deblender to recover first isolated and then blended sources, comparing against the official object catalogs to assess fidelity. This allows us to both gauge the initial performance of the model and to identify areas where the framework needs improvement. In future work we plan to explore the use of real data to train the model, through transfer learning or by fully re-training the networks and envision using space-based surveys for a joint-survey approach.


Arcelin, B., Doux, C., Aubourg, E., Roucelle, C., & LSST Dark Energy Science Collaboration (2021). Deblending galaxies with variational autoencoders: A joint multiband, multi-instrument approach. MNRAS, 500(1), 531-547.

Auteur

Lia Lammert (APC / CNRS)

Co-auteurs

Katherine Laliotis Anna Preto Cecile Roucelle Eric Aubourg

Documents de présentation

Aucun document.