Constraining cosmology with star-forming galaxies

PhD Program

Speaker

Sara Ortega Martínez

When

2026/09/23
12:00

Place

CFM (CSIC-EHU) auditorium, Donostia / San Sebastián

Add to calendar

iCal

PhD Thesis defense by Sara Ortega Martínez

Supervisor: Raúl Angulo (Ikerbasque Research Professor) and Sergio Contreras 

Cosmology

With new surveys mapping an unprecedented number of galaxies over the redshift range 0.6 < z < 2.5, we have an opportunity to better understand the composition and evolution of the Universe we live in, as well as the nature of gravity on cosmological scales. These new surveys (DESI, Euclid, Roman and PSF, among others) will target galaxies with strong emission lines, typically young, blue and star-forming galaxies (Emission Line Galaxies, ELG).

Within our current understanding of the Universe, galaxies form inside collapsed structures of dark matter, and it is possible to relate the galaxy and dark matter distribution thanks to the correlation between their formation and evolution. One approach to match galaxies to dark matter structures is population models, which paint galaxies into halos and subhalos based on their properties, and allow us to reproduce the spatial distribution and other features of the observed galaxy populations.

While scales above 20 h^-1Mpc are usually the main choice for constraining cosmological parameters, given the interplay between galaxy formation physics and cosmology at smaller scales, accessing these scales is crucial to maximise the information we can obtain from the data, which requires models that accurately describe galaxy clustering even in the inner regions of halos. In this thesis, divided into three publications, we develop the SHAMe-SF model to populate gravity-only simulations to reproduce the clustering of star-forming and emission-line galaxies, analyse the connection between them and their host dark matter (sub)halos, and constrain cosmology using galaxy clustering in the very non-linear regime.

Since we aimed to develop a physically motivated model within the subhalo abundance-matching framework, in the first publication, we used a hydrodynamical simulation to analyse the relationship between galaxies' star-formation rates and a wide range of subhalo properties. We used Random Forest, an ensemble machine learning regression method, to find the minimal set of subhalo properties that could recover the galaxy clustering of a star-formation rate-selected sample in the simulation. We found that we could select the most star-forming galaxies using a combination of the peak-over-redshift of the maximum circular velocity of subhalos, their host halo mass, and a concentration proxy. After selecting the properties, we parametrised the dependence of the simulation's star-formation rates, resulting in an 8-parameter model. We tested the model's flexibility in fitting the galaxy clustering of star-forming and colour-selected samples in an hydrodynamical simulation and in a semi-analytical model at various redshifts and number densities, and found that we could reproduce the clustering signal in all cases. In the second publication, we added three more parameters to the model, allowing greater flexibility for the most massive subhalos, which are not well constrained by galaxy clustering.

After fitting the galaxy clustering of a given sample, population models can be used to learn more about the connection between subhalos and the galaxies they host. In the second publication in the thesis, we fit the galaxy clustering between scales 0.1 < r < 30 h^-1Mpc for ELG data from the DESI One-Percent data release and for two ELG-like samples defined in a hydrodynamical simulation and a semi-analytical model. Aterwards, we analyse the predictions of the SHAMe-SF model for satellite abundance and distributions, host halo masses, assembly bias, and conformity. Compared to the assumptions and inferences of models analysing the same galaxies, we find similar average halo masses ⟨M_200,c⟩ ∼ 10^11.8 h⁻¹M⊙ when they are centrals, ⟨M_200,c⟩ ∼ 10^12.6 h⁻¹M⊙ when they are satellites), assembly bias (∼10%), and the number of satellites around central galaxies that are also in the sample (∼3%), but we tend to measure higher satellite fractions (∼30%). A combination of conformity and satellite anisotropy can explain the strong clustering signal at small scales.

Since SHAMe-SF was created with the intention of constraining cosmology, we developed and validated an inference pipeline in the third publication of the thesis. We trained an emulator (neural network-based interpolator) that predicts galaxy clustering for a given set of SHAMe-SF parameters, cosmological parameters, redshifts, and number densities, in a fraction of the time required to populate a simulation and compute the clustering statistics. We leveraged the differentiable nature of neural networks to sample the parameter space using a variant of Hamiltonian Monte Carlo, obtaining constraints on σ₈ and Ω_M for the One-Percent data release from DESI, and two samples defined on a hydrodynamical simulation (DESI-like and Euclid/Roman-like) by fitting galaxy clustering in the range 0.3 < r < 30 h^-1Mpc. Since the true cosmology of the simulation is known, we analysed the dependence of our constraints on the choice of clustering statistics and minimum scales, finding that small scales are necessary to avoid biased constraints on σ₈ due to degeneracies with one of the quenching parameters in SHAMe-SF.

Overall, we present the SHAMe-SF model and its capabilities, propose new directions including additional measurements and statistics, and leave the door open to investigate how it can be used in combination with other inference frameworks.