Project leads: Nicola R. Napolitano (University of Naples Federico II), Crescenzo Tortora (INAF-Osservatorio Astronomico di Capodimonte)
Point of contact: Crescenzo Tortora (crescenzo.tortora@inaf.it), Nicola R. Napolitano (nicolarosario.napolitano@unina.it)
Rubin project code: ITA-INA-S9
Relevant working groups: Galaxy Morphology
Project status: Active
The structural properties of galaxies—such as size, light concentration, ellipticity, and flux—are key observational probes for reconstructing galaxy formation and evolutionary pathways across cosmic time. Galaxy structure can be quantified visually, via non-parametric measurements, or by fitting analytical surface brightness models such as the 2D Sérsic profile. While machine learning classification models (e.g., identifying mergers versus non-mergers) are trained for specific single tasks and require re-training for new applications, quantitative structural parameters (magnitude, effective radius, Sérsic index, axis ratio, position angle) provide a universal, reusable framework for a wide array of scientific studies without needing to re-measure parameters.
The upcoming Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will yield deep imaging for billions of galaxies. This unprecedented volume presents two critical challenges for standard structural analysis pipelines:
Computational bottlenecks: Traditional parametric profile-fitting software packages (e.g., GALFIT, SourceXtractor++) rely on CPU-intensive iterative optimization routines. Processing individual galaxies takes standard tools anywhere from seconds to minutes, making them computationally prohibitive for billion-object LSST data streams.
Observational & atmospheric effects: Ground-based imaging quality is affected by point spread function (PSF) broadening, noise variations across filters/data releases, and crowded-field blending artifacts, all of which introduce systematic biases into structural measurements if not properly modeled alongside the image data.
To overcome these challenges, we utilize GAlaxy Light profile convolutional neural NETwork (GALNET, Li et al. 2022), a deep convolutional neural network designed to derive automated, unbiased, single- or multiple-component 2D Sérsic parameters directly from image cutouts and localized PSF models. Bypassing iterative profile optimization through deep multi-parameter regression allows the pipeline to perform rapid structural parameter estimation on massive galaxy samples.
We develop GALNET-LSST, a version of GALNET trained on LSST data (DP1, DP2, DR1, etc.), and apply it to new LSST datasets to derive structural parameters in the Rubin bands for billions of galaxies, enabling structural characterization at an unprecedented scale.
In Unni et al. (2026, in preparation), we present the first results from the training of GALNET to recover the structural parameters of single Sérsic profiles from galaxy images, using HSC and LSST DP1 data. The main results are summarized below:
High-fidelity parameter recovery: Validated on synthetic test benches (HSC and LSSTComCam data), GALNET accurately recovers true input parameters with near-perfect magnitude recovery, and low scatter in effective radius and Sérsic index.
Processing acceleration: Demonstrates massive computational efficiency gains over traditional profile-fitting software. GALNET processes a sample of 3,000 real galaxies in ∼ 30 seconds, compared to ∼ 30 minutes required by SourceXtractor++ on identical hardware.
Robust real-data performance: Benchmarked on HSC PDR3 and LSSTComCam DP1 observations, showing strong consistency with traditional tools while demonstrating high resistance to nearby blending artifacts and local background noise.
Proof-of-concept scaling relations: Successfully reproduced key galaxy physical relations in the Extended Chandra Deep Field South (ECDFS) up to z ∼ 1.2, including cosmic size growth (Re vs. absolute magnitude), color-redshift evolution, and structural separation between red passive spheroids and blue star-forming disks.
Future work. Ongoing work will focus on
Training GALNET on LSST DP2/DR1 data and applying it to the full DP2/DR1 survey footprints.
Extending GALNET to perform bulge–disk decomposition of galaxies.
Fine-tuning GALNET on real galaxy images to improve the modelling of complex morphologies, including structured and spiral galaxies.
The code developed during this project will be published and available on GitHub.
This work is complementary to the two projects, "Quantitative galaxy morphology with ML" led by Liza Sazonova and "Detection of morphological structures using CNNs" led by Vazquez-Mata, which focus on the morphological classification of galaxies