arXiv subjects

Serafina Di Gioia

Publications and source records attributed to Serafina Di Gioia.

5 recordsLinked to original sources

arXiv2026

Dark Matter profiles of "in silico" galaxies: deep learning inference

Machine learning has the potential to improve the reconstruction of the dark matter profile of galaxies with respect to traditional methods, like rotation curves. We demonstrate on the simulation suite Illustris-TNG that a steerable equivariant convolutional neural network (CNN) is able to infer the dark matter profiles within and around individual galaxies from photometric and interferometric data, improving on a standard CNN. Within the in silico environment of the simulations, our architecture is able to capture the dark matter distribution within galaxies without a parametrization of the profile. We perform an interpretability analysis to understand the internal mechanisms of the trained model and the most important data features used to estimate the dark matter profiles. The equivariant CNN recovers the dark matter profile of galaxies within the stellar mass range $[10^{10} - 10^{12} ]$ $M_{\odot}$ with excellent precision and accuracy: the mean squared error is reduced by a factor of ~ 3 from its value under the training distribution, demonstrating that the network has learnt from the data features. While this holds within the controlled 'in silico' environment of the simulation, we argue that few additional steps are needed before this method can be reliably applied to galaxies in the real field observations.

astro-ph.GA↗

arXiv2026

CORDEX-ML-Bench: A Benchmark for Data-Driven Regional Climate Downscaling -Experiment Design and Overview

Machine learning (ML) has emerged as a cost-effective approach to complement dynamical downscaling for producing high-resolution regional climate projections. However, the absence of standardised training and evaluation protocols, applied consistently across multiple domains, continues to hinder meaningful model intercomparison. We introduce CORDEX-ML-Bench, a benchmark aligned with CORDEX, which constitutes the first phase of a community initiative to advance data-driven downscaling toward operational readiness, and complement future dynamical downscaling efforts under CMIP7. The framework targets downscaled daily maximum temperature and precipitation to ~10 km resolution (20x increase) across three pilot regions; European Alps, New Zealand, and Southern Africa. Using a perfect-model experimental design, we evaluate 40 ML configurations developed independently, spanning traditional ML, convolutional U-Nets, vision transformers, graph neural networks, and generative models based on diffusion, flow matching, and generative adversarial networks. Models are trained under two experimental periods, an empirical-statistical downscaling pseudo-reality (historical period only) and Emulator (historical and future periods) -and are evaluated against a core set of metrics developed specifically for assessing downscaling skill. Generative models consistently outperform deterministic approaches for precipitation, better capturing fine-scale variability and extremes. For temperature, the generative advantage narrows and deterministic architectures remain competitive. Models trained solely on the historical period systematically underestimate future climate-change signals while those additionally trained on a future period perform better. These findings raise concerns about historically trained models widely used in an operational setting, underscoring the need for rigorous extrapolation testing.

physics.ao-ph↗

arXiv2026

Rotary Masked Autoencoders are Versatile Learners

Applying Transformers to irregular time-series typically requires specializations to their baseline architecture, which can result in additional computational overhead and increased method complexity. We present the Rotary Masked Autoencoder (RoMAE), which utilizes the popular Rotary Positional Embedding (RoPE) method for continuous positions. RoMAE is an extension to the Masked Autoencoder (MAE) that enables interpolation and representation learning with multidimensional continuous positional information while avoiding any time-series-specific architectural specializations. We showcase RoMAE's performance on a variety of modalities including irregular and multivariate time-series, images, and audio, demonstrating that RoMAE surpasses specialized time-series architectures on difficult datasets such as the DESC ELAsTiCC Challenge while maintaining MAE's usual performance across other modalities. In addition, we investigate RoMAE's ability to reconstruct the embedded continuous positions, demonstrating that including learned embeddings in the input sequence breaks RoPE's relative position property.

cs.LG↗

arXiv2025

On the effects of parameters on galaxy properties in CAMELS and the predictability of $Ω_{\rm m}$

Recent analyses of cosmological hydrodynamic simulations from CAMELS have shown that machine learning models can predict the parameter describing the total matter content of the universe, $Ω_{\rm m}$, from the features of a single galaxy. We investigate the statistical properties of two of these simulation suites, IllustrisTNG and ASTRID, confirming that $Ω_{\rm m}$ induces a strong displacement on the distribution of galaxy features. We also observe that most other parameters have little to no effect on the distribution, except for the stellar-feedback parameter $A_{SN1}$, which introduces some near-degeneracies that can be broken with specific features. These two properties explain the predictability of $Ω_{\rm m}$. We use Optimal Transport to further measure the effect of parameters on the distribution of galaxy properties, which is found to be consistent with physical expectations. However, we observe discrepancies between the two simulation suites, both in the effect of $Ω_{\rm m}$ on the galaxy properties and in the distributions themselves at identical parameter values. Thus, although $Ω_{\rm m}$'s signature can be easily detected within a given simulation suite using just a single galaxy, applying this result to real observational data may prove significantly more challenging.

astro-ph.CO↗

arXiv2020

Damped Lyman-$α$ absorbers and atomic hydrogen in galaxies: the view of the GAEA model

Using the GAEA semi-analytic model, we analyse the connection between Damped Lyman-$α$ systems (DLAs) and HI in galaxies. Our state-of-the-art semi-analytic model is tuned to reproduce the local galaxy HI mass function, and that also reproduces other important galaxy properties, including the galaxy mass - gas metallicity relation. To produce catalogs of simulated DLAs we throw $10^5$ random lines of sight in a composite simulated volume: dark matter haloes with log$(\frac{M_{200}}{ M_{\odot}}) \geq 11.5$ are extracted from the Millennium Simulation, while for $9.2 \leq \log(\frac{M_{200}}{ M_{\odot}})<11.5$ we use the Millennium II, and for $8 \leq \log(\frac{M_{200}}{M_{\odot}}) < 9.2$ a halo occupation distribution model. At $2 < z < 3$, where observational data are more accurate, our fiducial model predicts the correct shape of the column density distribution function, but its normalization falls short of the observations, with the discrepancy increasing at higher redshift. The agreement with observations is significantly improved increasing both the HI masses and the disk radii of model galaxies by a factor 2, as implemented 'a posteriori' in our $2M-2R$ model. In the redshift range of interest, haloes with $M_{200} \geq {10}^{11} M_{\odot}$ give the major contribution to $Ω_{\rm DLA}$, and the typical DLA host halo mass is $\sim {10}^{11} M _{\odot}$. The simulated DLA metallicity distribution is in relatively good agreement with observations, but our model predicts an excess of DLAs at low metallicities. Our results suggest possible improvements for the adopted modelling of the filtering mass and metal ejection in low-mass haloes.

astro-ph.GA↗