arXiv subjects

Natalí S. M. de Santi

Publications and source records attributed to Natalí S. M. de Santi.

10 recordsLinked to original sources

arXiv2026

Galaxy Phase-Space and Field-Level Cosmology: The Strength of Semi-Analytic Models

Semi-analytic models are a widely used approach to simulate galaxy properties within a cosmological framework, relying on simplified yet physically motivated prescriptions. They have also proven to be an efficient alternative for generating accurate galaxy catalogs, offering a faster and less computationally expensive option compared to full hydrodynamical simulations. In this paper, we demonstrate that using only galaxy $3$D positions and radial velocities, we can train a graph neural network coupled to a moment neural network to obtain a robust machine learning based model capable of estimating the matter density parameters, $Ω_{\rm m}$, with a precision of approximately 10%. The network is trained on ($25 h^{-1}$Mpc)$^3$ volumes of galaxy catalogs from L-Galaxies and can successfully extrapolate its predictions to other semi-analytic models (GAEA, SC-SAM, and Shark) and, more remarkably, to hydrodynamical simulations (Astrid, SIMBA, IllustrisTNG, and SWIFT-EAGLE). Our results show that the network is robust to variations in astrophysical and subgrid physics, cosmological and astrophysical parameters, and the different halo-profile treatments used across simulations. This suggests that the physical relationships encoded in the phase-space of semi-analytic models are largely independent of their specific physical prescriptions, reinforcing their potential as tools for the generation of realistic mock catalogs for cosmological parameter inference.

astro-ph.CO↗

arXiv2026

Cosmology with one galaxy: An analytic formula relating $Ω_{\rm m}$ with galaxy properties

Standard cosmological analyses typically treat galaxy formation and cosmological parameter inference as decoupled problems, relying on population-level statistics such as clustering, lensing, or halo abundances. However, classical studies of baryon fractions in massive galaxy clusters have long suggested that gravitationally bound systems may retain cosmological information through their baryonic content. Building on this insight, we present the first analytic and physically interpretable cosmological tracer that links the matter density parameter, $Ω_m$, directly to intrinsic galaxy-scale observables, demonstrating that cosmological information can be extracted from individual galaxies. Using symbolic regression applied to state-of-the-art hydrodynamical simulations from the CAMELS project, we identify a compact functional form that robustly recovers $Ω_m$ across multiple simulation suites (IllustrisTNG, ASTRID, SIMBA, and Swift-EAGLE), requiring only modest recalibration of a small number of coefficients. The resulting expression admits a transparent physical interpretation in terms of baryonic retention and enrichment efficiency regulated by gravitational potential depth, providing a clear explanation for why $Ω_m$ is locally encoded in galaxy properties. Our work establishes a direct, interpretable bridge between small-scale galaxy physics and large-scale cosmology, opening a complementary pathway to cosmological inference that bypasses traditional clustering-based statistics and enables new synergies between galaxy formation theory and precision cosmology.

astro-ph.CO↗

arXiv2026

Predicting galaxy bias using machine learning

Understanding how galaxies trace the underlying matter density field is essential for characterizing the influence of the large-scale structure on galaxy formation, being therefore a key ingredient in observational cosmology. This connection, commonly described through the galaxy bias, $b$, can be studied effectively using machine learning (ML) techniques, which offer strong predictive capabilities and can capture non-linear relationships. We aim to incorporate the linear bias parameter assigned to individual galaxies into a ML framework, quantify its dependence on various halo and environmental properties, and evaluate whether different algorithms can accurately predict this parameter and reproduce the scatter in several bias relations. We use data from the IllustrisTNG300 simulation, including the distance to different cosmic-web structures computed with DisPerSE. These data are complemented with an object-by-object estimator of the large-scale linear bias ($b_i$), providing the individual contribution of each galaxy to the bias of the entire population. Our ML framework uses three models to predict $b_i$: a Random Forest Regressor, a Neural Network and a probabilistic method (Normalizing Flows). We recover the full hierarchy of galaxy bias dependencies, showing that the most informative features are the overdensities, particularly $δ_8$, followed by the distances to cosmic-web structures and selected internal halo properties, most notably $z_{1/2}$. We also demonstrate that Normalizing Flows clearly outperform deterministic methods in predicting galaxy bias, including its joint distributions with galaxy properties, owing to their ability to capture the intrinsic variance associated with the stochastic nature of the matter-halo-galaxy connection. Our ML framework provides a foundation for future efforts to measure individual bias with upcoming spectroscopic surveys.

astro-ph.CO↗

arXiv2025

Field-level simulation-based inference with galaxy catalogs: the impact of systematic effects

It has been recently shown that a powerful way to constrain cosmological parameters from galaxy redshift surveys is to train graph neural networks to perform field-level likelihood-free inference without imposing cuts on scale. In particular, de Santi et al. (2023) developed models that could accurately infer the value of $Ω_{\rm m}$ from catalogs that only contain the positions and radial velocities of galaxies that are robust to uncertainties in astrophysics and subgrid models. However, observations are affected by many effects, including 1) masking, 2) uncertainties in peculiar velocities and radial distances, and 3) different galaxy selections. Moreover, observations only allow us to measure redshift, intertwining galaxies' radial positions and velocities. In this paper we train and test our models on galaxy catalogs, created from thousands of state-of-the-art hydrodynamic simulations run with different codes from the CAMELS project, that incorporate these observational effects. We find that, although the presence of these effects degrades the precision and accuracy of the models, and increases the fraction of catalogs where the model breaks down, the fraction of galaxy catalogs where the model performs well is over 90 %, demonstrating the potential of these models to constrain cosmological parameters even when applied to real data.

astro-ph.CO↗

arXiv2024

Exploring the halo-galaxy connection with probabilistic approaches

The connection between galaxies and dark matter halos encompasses a range of processes and play a pivotal role in our understanding of galaxy formation and evolution. Traditionally, this link has been established through physical or empirical models. Machine learning techniques are adaptable tools that handle high-dimensional data and grasp associations between numerous attributes. In particular, probabilistic models capture the stochasticity inherent to these complex relations. We compare different probabilistic machine learning methods to model the uncertainty in the halo-galaxy connection and efficiently generate galaxy catalogs that faithfully resemble the reference sample by predicting joint distributions of central galaxy properties conditioned to their host halo features. The analysis is based on the IllustrisTNG300 simulation. The methods model the distributions in different ways. We compare a multilayer perceptron that predicts the parameters of a multivariate Gaussian distribution, a multilayer perceptron classifier, and the method of normalizing flows. The classifier predicts the parameters of a Categorical distribution, which are defined in a high-dimensional parameter space through a Voronoi cell-based hierarchical scheme. We evaluate the model's performances under various sample selections based on halo properties. The three methods exhibit comparable results, with normalizing flows showing the best performance in most scenarios. The models reproduce the main features of galaxy properties distributions with high-fidelity and reproduce the results obtained with traditional, deterministic, estimators. Our results also indicate that different halos and galaxy populations are subject to varying degrees of stochasticity, which has relevant implications for studies of large-scale structure.

astro-ph.CO↗

arXiv2023

Robust Field-level Likelihood-free Inference with Galaxies

We train graph neural networks to perform field-level likelihood-free inference using galaxy catalogs from state-of-the-art hydrodynamic simulations of the CAMELS project. Our models are rotational, translational, and permutation invariant and do not impose any cut on scale. From galaxy catalogs that only contain $3$D positions and radial velocities of $\sim 1, 000$ galaxies in tiny $(25~h^{-1}{\rm Mpc})^3$ volumes our models can infer the value of $Ω_{\rm m}$ with approximately $12$ % precision. More importantly, by testing the models on galaxy catalogs from thousands of hydrodynamic simulations, each having a different efficiency of supernova and AGN feedback, run with five different codes and subgrid models - IllustrisTNG, SIMBA, Astrid, Magneticum, SWIFT-EAGLE -, we find that our models are robust to changes in astrophysics, subgrid physics, and subhalo/galaxy finder. Furthermore, we test our models on $1,024$ simulations that cover a vast region in parameter space - variations in $5$ cosmological and $23$ astrophysical parameters - finding that the model extrapolates really well. Our results indicate that the key to building a robust model is the use of both galaxy positions and velocities, suggesting that the network have likely learned an underlying physical relation that does not depend on galaxy formation and is valid on scales larger than $\sim10~h^{-1}{\rm kpc}$.

astro-ph.CO↗

arXiv2023

The CAMELS project: Expanding the galaxy formation model space with new ASTRID and 28-parameter TNG and SIMBA suites

We present CAMELS-ASTRID, the third suite of hydrodynamical simulations in the Cosmology and Astrophysics with MachinE Learning (CAMELS) project, along with new simulation sets that extend the model parameter space based on the previous frameworks of CAMELS-TNG and CAMELS-SIMBA, to provide broader training sets and testing grounds for machine-learning algorithms designed for cosmological studies. CAMELS-ASTRID employs the galaxy formation model following the ASTRID simulation and contains 2,124 hydrodynamic simulation runs that vary 3 cosmological parameters ($Ω_m$, $σ_8$, $Ω_b$) and 4 parameters controlling stellar and AGN feedback. Compared to the existing TNG and SIMBA simulation suites in CAMELS, the fiducial model of ASTRID features the mildest AGN feedback and predicts the least baryonic effect on the matter power spectrum. The training set of ASTRID covers a broader variation in the galaxy populations and the baryonic impact on the matter power spectrum compared to its TNG and SIMBA counterparts, which can make machine-learning models trained on the ASTRID suite exhibit better extrapolation performance when tested on other hydrodynamic simulation sets. We also introduce extension simulation sets in CAMELS that widely explore 28 parameters in the TNG and SIMBA models, demonstrating the enormity of the overall galaxy formation model parameter space and the complex non-linear interplay between cosmology and astrophysical processes. With the new simulation suites, we show that building robust machine-learning models favors training and testing on the largest possible diversity of galaxy formation models. We also demonstrate that it is possible to train accurate neural networks to infer cosmological parameters using the high-dimensional TNG-SB28 simulation set.

astro-ph.CO↗

arXiv2023

High-fidelity reproduction of central galaxy joint distributions with Neural Networks

The relationship between galaxies and haloes is central to the description of galaxy formation, and a fundamental step towards extracting precise cosmological information from galaxy maps. However, this connection involves several complex processes that are interconnected. Machine Learning methods are flexible tools that can learn complex correlations between a large number of features, but are traditionally designed as deterministic estimators. In this work, we use the IllustrisTNG300-1 simulation and apply neural networks in a binning classification scheme to predict probability distributions of central galaxy properties, namely stellar mass, colour, specific star formation rate, and radius, using as input features the halo mass, concentration, spin, age, and the overdensity on a scale of 3 $h^{-1}$ Mpc. The model captures the intrinsic scatter in the relation between halo and galaxy properties, and can thus be used to quantify the uncertainties related to the stochasticity of the galaxy properties with respect to the halo properties. In particular, with our proposed method, one can define and accurately reproduce the properties of the different galaxy populations in great detail. We demonstrate the power of this tool by directly comparing traditional single-point estimators and the predicted joint probability distributions, and also by computing the power spectrum of a large number of tracers defined on the basis of the predicted colour-stellar mass diagram. We show that the neural networks reproduce clustering statistics of the individual galaxy populations with excellent precision and accuracy.

astro-ph.CO↗

arXiv2022

Improving cosmological covariance matrices with machine learning

Cosmological covariance matrices are fundamental for parameter inference, since they are responsible for propagating uncertainties from the data down to the model parameters. However, when data vectors are large, in order to estimate accurate and precise matrices we need huge numbers of observations, or rather costly simulations - neither of which may be viable. In this work we propose a machine learning approach to alleviate this problem in the context of the matrices used in the study of large-scale structure. With only a small amount of data (matrices built with samples of 50-200 halo power spectra) we are able to provide significantly improved matrices, which are almost indistinguishable from the ones built from much larger samples (thousands of spectra). In order to perform this task we trained convolutional neural networks to denoise the matrices, using in the training process a data set made up entirely of spectra extracted from simple, inexpensive halo simulations (mocks). We then show that the method not only removes the noise in the matrices of the cheap simulation, but it is also able to successfully denoise the matrices of halo power spectra from N-body simulations. We compare the denoised to the other matrices using several metrics, and in all of them they score better, without any signs of spurious artifacts. With the help of the Wishart distribution we derive an analytical extrapolation for the effective sample augmentation allowed by the denoiser. Finally, we show that, by using the denoised matrices, the cosmological parameters can be recovered with nearly the same accuracy as when using matrices built with a sample of 30,000 spectra in the case of the cheap simulations, and with 15,000 spectra in the case of the N-body simulations. Of particular interest is the bias in the Hubble parameter $H_0$, which was significantly reduced after applying the denoiser.

astro-ph.CO↗

arXiv2022

Mimicking the halo-galaxy connection using machine learning

Elucidating the connection between the properties of galaxies and the properties of their hosting haloes is a key element in galaxy formation. When the spatial distribution of objects is also taken under consideration, it becomes very relevant for cosmological measurements. In this paper, we use machine learning techniques to analyse these intricate relations in the IllustrisTNG300 magnetohydrodynamical simulation, predicting baryonic properties from halo properties. We employ four different algorithms: extremely randomized trees, K-nearest neighbours, light gradient boosting machine, and neural networks, along with a unique and powerful combination of the results from all four approaches. Overall, the different algorithms produce consistent results in terms of predicting galaxy properties from a set of input halo properties that include halo mass, concentration, spin, and halo overdensity. For stellar mass, the Pearson correlation coefficient is 0.98, dropping down to 0.7-0.8 for specific star formation rate (sSFR), colour, and size. In addition, we apply, for the first time in this context, an existing data augmentation method, synthetic minority over-sampling technique for regression with Gaussian noise (SMOGN), designed to alleviate the problem of imbalanced data sets, showing that it improves the overall shape of the predicted distributions and the scatter in the halo-galaxy relations. We also demonstrate that our predictions are good enough to reproduce the power spectra of multiple galaxy populations, defined in terms of stellar mass, sSFR, colour, and size with high accuracy. Our results align with previous reports suggesting that certain galaxy properties cannot be reproduced using halo features alone.

astro-ph.GA↗