Choose timezone
Your profile timezone:
Third Technical Meeting Spoke 3
→
Europe/Rome
Perugia
Perugia
Description
The Third General Technical workshop of the Spoke 3 ASTROPHYSICS & COSMOS OBSERVATIONS.
The workshop is dedicated to discuss the technical activites of the Spoke3 partners. The workshop will be hosted by INFN Perugia.
Speakers are requested to download the following template and use it to organize their presentation: Presentation Template
ON-LINE link participation is POSSIBLE: Zoom link
Address
Hotel Giò Jazz & Wine
Via Ruggero D'Andreotto, 19, Perugia, PG, Italia
Tel: +39 075 5731100
https://www.hotelgio.it
-
-
IntroductionConvener: Pasquale Lubrano (INFN Perugia)
- 1
-
2
Spoke3 Status ReportSpeaker: Ugo Becciani (Istituto Nazionale di Astrofisica (INAF))
-
3
Technical ReportSpeakers: Claudio Gheller (Istituto Nazionale di Astrofisica (INAF)), Giuliano Taffoni (Istituto Nazionale di Astrofisica (INAF))
-
WP5Conveners: Dario Gasparrini (Istituto Nazionale di Fisica Nucleare sez. di Roma Tor Vergata), Diego Ciangottini (INFN)
-
4
WP5 OverviewSpeaker: Dario Gasparrini (Istituto Nazionale di Fisica Nucleare sez. di Roma Tor Vergata)
-
5
The scientific hub infrastructure: from datalake to HPC computingSpeaker: Diego Ciangottini (INFN)
-
4
-
Coffee break
-
Wp4: WP4Conveners: Cristina Knapic (Istituto Nazionale di Astrofisica (INAF)), Deborah Busonero (Istituto Nazionale di Astrofisica (INAF))
-
6
WP4 activities updateSpeaker: Cristina Knapic (Istituto Nazionale di Astrofisica (INAF))
-
7
Spoke3 Archive InfrastructureSpeaker: Massimo Costantini (Istituto Nazionale di Astrofisica (INAF))
-
8
Spoke3: Simulation Runs Discovery ServiceSpeaker: Hendrik Heinl (Istituto Nazionale di Astrofisica (INAF))
-
9
The Gaia use case in Spoke3Speaker: Sara Gelsumini (Istituto Nazionale di Astrofisica (INAF))
-
6
-
-
-
Key Science ProjectsConvener: Claudio Gheller (Istituto Nazionale di Astrofisica (INAF))
-
10
Porting of VisIVO Visual Analytics on HPC-type computing infrastructure
We present the latest developments of VisIVO Visual Analytics, an interactive visualization tool that is evolving to meet the needs of the SKA and similar large-scale astrophysical projects. Originally designed as a desktop application, VisIVO is transitioning to a client-server architecture, allowing it to efficiently run remote visualization pipelines on remote servers. This shift also enable the ability to distribute computational workloads across multiple nodes in high-performance computing (HPC) clusters for parallel visualization pipelines.
Speaker: Giuseppe Tudisco (Istituto Nazionale di Astrofisica (INAF)) -
11
KSP - OPAL: Update on the production phaseSpeaker: Danai Polychroni (Istituto Nazionale di Astrofisica (INAF))
-
10
-
WP1-WP2Convener: Stefano Della Torre (INFN Sez. Milano-Bicocca)
-
12
RICK - HPC and GPU development for radio astronomy data reductionSpeaker: Emanuele De Rubeis (Istituto Nazionale di Astrofisica (INAF))
-
13
OpenGadget3: current status and perspectivesSpeaker: Milena Valentini (University of Trieste)
-
14
OpenGadget3: an improved version of black hole dynamicsSpeaker: Alice Damiano (Istituto Nazionale di Astrofisica (INAF))
-
12
-
Coffee break
-
WP1-WP2Convener: Stefano Della Torre (INFN Sez. Milano-Bicocca)
-
15
BrahMap: Map-making for future CMB experimentsSpeaker: Mr Avinash Anand (University of Rome "Tor Vergata")
-
16
GalaPy - the Spectral modelling tool for galaxies in python.Speaker: Tommaso Ronconi (Scuola Internazionale Superiore di Studi Avanzati (SISSA), Istituto Nazionale di Astrofisica (INAF))
-
17
GUIBRUSH(R)-Graphic User Interface for Bayesian Retrieval Using Spectroscopy at High ResolutionSpeaker: Gloria Guilluy (Istituto Nazionale di Astrofisica (INAF))
-
18
Parallelizing the Mercury-Arxes code with OpenACC: an updateSpeaker: Paolo Matteo Simonetti (Istituto Nazionale di Astrofisica (INAF))
-
19
GPU porting of FFT and collapse time calculations within PINOCCHIOSpeaker: Marius Daniel Lepinzan (Istituto Nazionale di Astrofisica (INAF))
-
20
PRESTO Jerksearch GPU porting - updates and developemnt statusSpeaker: Rouhin Nag (INAF OAC)
-
21
High Performance StingraySpeaker: Dr Eleonora Veronica Lai (Istituto Nazionale di Astrofisica (INAF))
-
15
-
Coffee Break
-
WP1-WP2Conveners: Mario Spera (Istituto Nazionale di Astrofisica (INAF)), Stefano Della Torre (INFN Sez. Milano-Bicocca)
-
22
NP_TMcode project status reportSpeaker: Dr Giovanni La Mura (INAF - Osservatorio Astronomico di Cagliari)
-
23
Sparse spectral-imaging and component separation algorithmsSpeaker: Federico De Luca (Università degli Studi di Roma Tor Vergata)
-
24
gPLUTO: a multi-physics code for astrophysical plasma.Speaker: Andrea Mignone (UNITO)
-
25
TURBO 3D: Complex flows and Lagrangian observables.Speaker: Victor de Jesus Valadão (UNITO)
-
26
CAMB_GPU: porting of a cosmological boltzmann solverSpeaker: Paolo Campeti (INFN Sezione di Ferrara, ICSC)
-
27
StratLearn updatesSpeaker: Chiara Moretti (Istituto Nazionale di Astrofisica (INAF))
-
22
-
Coffee break
-
-
-
WP3Conveners: Fabio Gargano (INFN Bari), Fabio Roberto Vitello
-
28
From few to many maps: A fast map-level emulator for extreme augmentation of CMB systematics datasets
We introduce a novel, fast, and efficient generative model built upon scattering covariances, the most recent iteration of the scattering transforms statistics. This model is designed to augment by several orders of magnitude the number of map simulations in datasets
of computationally expensive Cosmic Microwave Background (CMB) instrumental systematics simulations, including their non-Gaussian and inhomogeneous features. Unlike conventional neural network-based algorithms, this generative model requires only
a minimal number of training samples, making it highly compatible with the computational constraints of typical CMB simulation campaigns. While our primary focus is on spherical data, the framework is inherently versatile and readily applicable to 1D and 2D planar data, leveraging the localized nature of scattering statistics. We validate the method using realistic simulations of CMB systematics, which are particularly challenging to emulate, and perform extensive statistical tests to confirm its ability to produce new statistically independent approximate realizations. Remarkably, even when trained on as few as 10 simulations, the emulator closely reproduces key summary statistics—including the angular power spectrum, scattering coefficients, and Minkowski functionals—and provides pixel covariance estimates with substantially reduced sample noise compared to those obtained without augmentation. The proposed approach has the potential to shift the paradigm in simulation campaign design. Instead of producing large numbers of low- or medium-accuracy simulations, future pipelines can focus on generating a few high-accuracy simulations that are then efficiently augmented using such generative model. This promises significant benefits not only for current and forthcoming cosmological surveys such as Planck, LiteBIRD, Simons Observatory, CMB-S4, Euclid and Rubin-LSST, but also for diverse fields including oceanography and climate science. We make both the general framework for scattering transform statistics HealpixML and the emulator CMBSCAT available to the community.Speaker: Paolo Campeti (INFN Sezione di Ferrara, ICSC) -
29
Updates on Radio U-Net: application to the full LoTSS survey
In this talk, I will present the latest developments in the Radio U-Net pipeline. Radio U-Net is a convolutional neural network derived from the U-Net architecture, specifically designed to perform rapid and automated segmentation of diffuse radio emission in extensive low-frequency surveys. Its application to the LOFAR Two Metre Sky Survey (LoTSS) has demonstrated high accuracy: when the segmentation maps are cross-matched with galaxy cluster catalogs, the network successfully recovered 83% of clusters exhibiting diffuse emission in a balanced sample of 246 clusters. These promising results have motivated the application of Radio U-Net to the entire LoTSS survey, which is scheduled for release in Summer 2025. We are currently refining the automatic pipeline with the ambitious goal of detecting diffuse radio emissions across a sample of approximately 3,000 galaxy clusters. I will show the achieved results and forecast the nextfinal steps to conclude our work. I will show the achieved results and forecast the final steps to conclude our work.
Speaker: Chiara Stuardi (Istituto Nazionale di Astrofisica (INAF)) -
30
Multi-Tracer Inference of the Epoch of Reionisation and Cosmic Dawn
Understanding the epochs of cosmic dawn and reionisation requires us to leverage multi-wavelength and multi-tracer observations, with each dataset providing a complimentary piece of the puzzle. To interpret such data, we update the public simulation code, 21cmFASTv4, to include a discrete source model based on stochastic sampling of conditional mass functions and semi-empirical galaxy relations.
We demonstrate that our new galaxy model is flexible enough to characterize very different predictions from hydrodynamic cosmological simulations of high-redshift galaxies.
Combining a discrete galaxy population with approximate, efficient radiative transfer
allows us to self-consistently forward-model galaxy surveys, line intensity maps (LIMs), and observations of the intergalactic medium (IGM).Not only does each observable probe different scales and physical processes, but cross-correlation will maximise the information gained from each measurement by probing the galaxy-IGM connection at high-redshift. Scatter in galaxy properties can be constrained using UV luminosity functions and/or 21cm power spectra, especially if astrophysical scatter is higher than expected (as might be needed to explain recent JWST observations).
Our modelling pipeline is both flexible and computationally efficient, facilitating high-dimensional, multi-tracer, field-level Bayesian inference of cosmology and astrophysics during the first billion years.Speaker: James Davies (Scuola Normale Superiore) -
31
Detecting the cross-correlation between 21cm signal and line intensity mapping during reionization
The distribution of neutral hydrogen mapped by SKA-Low through the 21 cm signal provides a unique and powerful probe of the cosmic reionization process. However, this method faces significant challenges due to severe foreground contamination, which can exceed the cosmological signal by up to five orders of magnitude.
Cross-correlating 21 cm and line intensity maps such as [CII] and CO offers a promising strategy to mitigate foregrounds and other systematics, as the dominant contaminants in these two surveys are largely uncorrelated. Detecting the cross power spectrum would verify initial claims of a detection with SKA. Moreover LIM and 21cm synergies very nicely as [CII] and CO lines originate from the interstellar medium while cosmic 21cm is sourced from the intergalactic medium. Together, they draw a complete view of the reionization process in the high-redshift Universe.
We develop a pipeline for simultaneously forward modeling 21cm, LIMs as their cross-correlation. We demonstrate that high signal to noise detection of 21cm-[CII] is possible with SKA-low AA* combined with LIM from several upcoming telescopes. Our study highlights the advantages of multi-tracer approaches in overcoming observational challenges and giving more physical insight into the epoch of reionization.Speaker: Yilong Zhang (Scuola Normale Superiore (Pisa)) -
32
Anomaly Detection with Machine Learning Techniques on the Fermi ACD Time Series DataSpeaker: Andrea Adelfio (INFN Perugia)
-
33
Bayesian inference for the nHz SGWB in PTA data analysis with ML
Complex inference tasks such as those in Pulsar Timing Array (PTA) data analysis rely on Bayesian frameworks. The high-dimensional parameters space and the strong interdependencies among astrophysical, pulsar noise and nuisance parameters introduce significant challenges. We address two of them. The first focuses on speeding up the existing code for Bayesian inference by using NessAI, a nested sampling algorithm for Bayesian Inference that makes use of a ML algorithm designed for applications where likelihood is computationally expensive, as is the case for PTA. I will present the improvements obtained for the analysis of the 25 pulsars of EPTADR2 in the CURN model. The second, in collaboration with Koexai s.r.l., focuses on reducing the dependence of the posterior for the parameters of the nHz SGWB on the choice of the prior of the nuisance parameters by means of a ML-guided reparametrization in the parameters space.
Speaker: Eleonora Villa (Istituto Nazionale di Astrofisica (INAF)) -
34
Advanced Tracking Analysis in Space Experiments with Graph Neural Networks
The integration of advanced Artificial Intelligence (AI) techniques into astroparticle experiments represents a transformative step for data analysis and experimental design. With the increasing complexity of space missions, the adoption of AI technologies becomes crucial for optimizing performance and achieving robust scientific outcomes.
This study focuses on the development of innovative AI-driven algorithms for track reconstruction, leveraging the potential of Graph Neural Networks (GNNs). GNNs, as a branch of geometric deep learning, are particularly well-suited for modeling the intrinsic structure of tracking detectors, where nodes correspond to energy deposits ('hits') and edges to their potential connections. These networks enable the development of targeted approaches for tasks such as node classification, link prediction, and graph classification, addressing the specific challenges of space-based experiments.
A critical obstacle in tracking systems for space experiments is the high-noise environment. This noise, characterized by hits originating from backscattering particles from the calorimeter and by electronic noise, significantly complicates the accurate identification of the primary particle trajectory.
To overcome this challenge, we propose an innovative GNN-based approach for node-level classification, specifically designed to distinguish signal hits (belonging to the primary particle track) from noise hits (including backscattering and electronic noise).
This algorithm effectively identifies the hits constituting the primary track within a noisy environment, thus facilitating the subsequent accurate retrieval of the particle's track parameters.
We will present updates on our work, aimed at demonstrating the potential of this GNN-based tracking algorithm applied to simulated HERD data.
Speaker: Federica Cuna (INFN-BARI) -
35
Machine Learning Techniques for Shower Discrimination in Space-Based Calorimeters for Cosmic Ray Studies
We present a machine learning approach for discriminating electromagnetic and hadronic showers in simulated data from a high-granularity LYSO space-based calorimeter. Two feature sets were explored: Vision Transformers were trained on the spatial coordinates and deposited energy of activated pixels, while XGBoost classifiers used reconstructed variables such as lateral and transverse moments and longitudinal profiles. The results show that both methods offer complementary insights, highlighting the potential of combining raw and physics-informed features for improved shower classification in cosmic ray studies.
Speaker: Maria Bossa (INFN)
-
28
-
Coffee break
-
WP3Conveners: Fabio Gargano (INFN Bari), Fabio Roberto Vitello
-
36
From cosmic rays to noise characterization in the microwave detector experiment
Simulating the effect of the Cosmic Rays (CR) on the signal of the High Frequency Telescope (HFT) of microwave detector experiment from space implies a computational expensive chain of Monte Carlo (MC) simulations. It started with the CR spectra at L2 mission location, then we propagated them into the detector materials, extract the hits deposited energies on the sensible area, propagate the heat in the material and convert them in the Transition Edge Sensor (TES) bolometers response. In this talk, we will illustrate the process of generating Time Ordered Data (TOD) and their features. The samples have to be representative of the 3 years mission acquisition period. So here we placed us with the Machine Learning (ML) CRAB code to perform data augmentation. The TODs are expanded with a convolutional Generative Adversarial Network (GAN), tuned on the MC simulated noise in HFT.
Speaker: Giovanni Cavallotto (INFN Milano Bicocca) -
37
Progress report on the use of old stellar tracers to constrain the early formation of the Galactic spheroid
The main aim of this project is to use RR Lyrae (RRL) variables to investigate the Milky Way’s early formation and evolution. We will provide the largest spectroscopic dataset for RRLs, using a mix of proprietary and publicly available spectroscopic catalogs. We provide accurate measurements of radial velocities and metallicities via the Delta-S method by using single epoch spectra. Moreover, we also investigate their orbital properties using kinematic parameters provided by the astrometric satellite Gaia. The chemo–dynamical properties of old (t>= 10 Gyrs) stellar tracers will allow us to investigate the early formation and evolution of the Galactic spheroid.
Speaker: Karina Baeza Villagra (Istituto Nazionale di Astrofisica (INAF)) -
38
Symmetric solutions for the N-body problem: a computational approachSpeaker: Irene De Blasi (UNITO)
-
36
-
Innovation Grants
-
39
Data Quality, ATS & Fraud Detection status ReportSpeaker: Dr Riccardo Crupi (Istituto Nazionale di Astrofisica (INAF))
-
40
Data Quality, ATS & Fraud detection
We will present recent advancements in our work on Data Quality, ATS, and Fraud Detection, developed in collaboration with Banca Intesa Sanpaolo. Specifically, we describe two complementary approaches. The first builds on our previous work combining Feature Engineering, Self-Organizing Maps (SOM), and Isolation Forest (IF). In this case, we focus on analyzing the behavior of specific instances across different time windows. The second approach investigates the use of autoencoders to identify potential outliers.
We will also briefly illustrate how the same methodology (Feature Engineering, SOM, and IF) can be applied to the Astrophysics domain, where it enables the identification of a relatively pure sample of Active Galactic Nuclei (AGN) among galaxies and stars.
Moreover, we introduce our fraud detection strategy, which uses Random Forest to identify potentially fraudulent events, and applies the K-Nearest Neighbors (KNN) algorithm at a user-level to identify anomalous operations by comparing each transaction to the user's typical behavior.
Finally, given the good results obtain in the high energy astrophysics, it is also presented briefly FoCUS, an evolution of cusum, and a possible application on the Banking data.Speaker: Ylenia Maruccia (Istituto Nazionale di Astrofisica (INAF))
-
39
-
WP3
-
41
Fast and Automated Characterization of Core-Collapse Supernovae with Machine Learning and HPC-Driven Modeling
The growing number of observed supernovae in current and upcoming surveys demands scalable and automated methods for their physical characterization. In this framework, we developed an AI-based pipeline to reconstruct bolometric light curves from multi-band photometric observations and infer physical parameters such as ejecta mass, explosion energy, and Ni-56 yield. Real data observation from surveys like ZTF, ATLAS have been analyzed through custom tools for data preprocessing, filtering, and quality assessment. Synthetic datasets—generated from semi-analytical and numerical codes run on the Galileo100 HPC cluster—have been used to train and validate the machine learning models. This approach enables rapid, consistent analysis across H-rich SNe and observational conditions. We also explored early applications of this framework to recent supernovae of interest, demonstrating its potential for real-time transient analysis and cross-matching with neutrino and gamma-ray observations.
Speaker: Stefano Pio Cosentino (Istituto Nazionale di Astrofisica (INAF))
-
41
-
13:00
Lunch
-
Innovation GrantsConvener: Claudio Gheller (Istituto Nazionale di Astrofisica (INAF))
-
42
The Hammon ProjectSpeaker: Antonio Tirri (Leithà)
-
43
HaMMon: Segmentation and Classification of aerial imagesSpeaker: Nicoletta Sanvitale (Istituto Nazionale di Astrofisica (INAF) - Istituto di Radioastronomia)
-
44
HaMMon: Automated Photogrammetric Workflow for Environmental Digital Twin Generation and Hazard Assessment
This work reports recent advancements in the development of an automated photogrammetric pipeline for generating 3D geospatial digital twins, Data Visualization and Artificial Intelligence technologies, aimed at environmental monitoring and hazard assessment.
For interactive 3D visualization and analysis, the pipeline integrates dense 3D reconstructions into the CesiumJS web environment and includes an automated point-cloud classification tool that uses image masks to enrich the models with semantic labels.
Recent efforts have also focused on restructuring the workflow to ensure greater automation, modularity, and reproducibility, facilitating its integration within a Science Gateway framework for broader accessibility and scalability.Speaker: Leonardo Pelonero (Istituto Nazionale di Astrofisica (INAF)) -
45
Serial Code Porting on HPC & Quantum ComputingSpeaker: Manuele Innocenti (SOGEI)
-
46
Interoperable Data Lake (IDL)Speaker: Cristina Knapic (Istituto Nazionale di Astrofisica (INAF))
-
47
IDL: INAF ContributionSpeaker: Giacomo Coran (Istituto Nazionale di Astrofisica (INAF))
-
42
-
Coffee Break
-
Innovation Grants
-
48
Status of IGUC projectSpeakers: Deborah Busonero (Istituto Nazionale di Astrofisica (INAF)), Paolo Giacomazzi (CherryData)
-
48
-
Bandi a Cascata
-
49
CATMAD: Synthesis of nonthermal emission from 3-D MHD simulations of astrophysical plasmas
CATMAD is a project aimed at developing an advanced diagnostic tool for the analysis of nonthermal emission stemming from supernova remnants and astrophysical plasmas.
Through post-processing of state-of-the-art 3-D MHD simulations of supernova remnants, we synthesise the multi-wavelength emission (from the radio band to gamma-rays) to get an accurate comparison with actual observations. We here present our preliminary results, showing our module for the synthesis of hadronic gamma-ray emission. The module incorporates the time-dependent acceleration of protons at the shock front, their escape during the evolution of the remnant, and the inelastic collisions with ambient protons in the post-shock material. We provide an application of our diagnostic tool to the Galactic supernova remnant IC443, by showing that we can self-consistently reproduce its gamma-ray emission and obtain important insight on the efficiency of the proton acceleration process, and on the cosmic-ray energy and spectral shape.Speaker: Fabrizio Bocchino (Istituto Nazionale di Astrofisica (INAF)) -
50
PAGUSCI - Porting of the CAMB code to GPU
In the presentation we will outline the essential stages required to profile and port the CAMB code onto a GPU architecture. The process encompasses detailed performance analysis of the existing CAMB code, identifying bottlenecks, and strategically adapting code segments for GPU execution. In this presentation, we will primarily focus on the code modifications necessary to adapt the CAMB code for the NVIDIA compiler, aiming to achieve the first functional GPU-compatible version. Subsequent to this initial porting, we will delve into optimization strategies to enhance the performance of the GPU-enabled code.
Speaker: Loriano Storchi (Chieti University) -
51
ASTRAI
We present ASTRAI, an end-to-end framework that accelerates Type II supernova (SN) research by integrating physically motivated models with generative diffusion techniques to produce a high-fidelity synthetic dataset for deep learning (DL)–based SN characterization. First, we aggregate and homogenize multi–survey photometric and spectroscopic observations from Carnegie, WISeREP, Zenodo, and LSST precursor programs, applying bolometric corrections and Gaussian-process interpolations to construct a benchmark dataset of real light curves. Building upon semi-analytical hydrodynamic and ejecta circumstellar medium (CSM) interaction models, we generate synthetic SN observables spanning total explosion energy, ejected mass, progenitor radius, radioactive yields, and CSM properties.
To address the irregular time sampling and incomplete labeling of real data, we develop a two-stage Generative AI pipeline: (1) a diffusion-based Light Curve Generator (LCGen) that learns the manifold of physically consistent SN evolutions conditioned on progenitor parameters; and (2) a Physical Parameter Regressor (PPR) that inverts the LCGen outputs back to progenitor properties. By jointly optimizing mean-squared losses and reconstruction error of the full cycle in both observable and parameter spaces, our approach ensures realism, diversity, and physical interpretability.
We evaluate ASTRAI on held-out surveys, demonstrating that DL models pre-trained on our synthetic dataset and fine-tuned on limited real observations reduce median parameter errors by half compared to classical Gaussian-process methods, while inference times drop from minutes to milliseconds per object. Ablation studies reveal that diffusion sampling contributes an improvement in light-curve fidelity over standard variational autoencoders, particularly in phases with sparse data. Our results showcase the potential of combining diffusion-driven data augmentation with physics-aware DL architectures to meet the demands of next-generation surveys such as LSST, enabling rapid, automated, and accurate characterization of supernovae.
Speaker: Vincenzo Del Zoppo (KOEXAI)
-
49
-
Social Dinner
-
-
-
Bandi a CascataConvener: Ugo Becciani
-
52
SPECTRA & GRAIL
The SPECTRA AI project addresses the challenge of identifying and classifying transient gamma-ray emissions in Fermi LAT observations using advanced artificial intelligence methods. Transient phenomena such as gamma-ray bursts and flares exhibit complex spatio-temporal dynamics, necessitating robust pattern-recognition frameworks.
The GRAIL project (Gamma-Ray Imaging with Deep Learning) aims to develop a novel approach for the analysis of calorimetric data in space by integrating artificial intelligence directly onboard satellites. Space-based calorimeters generate massive volumes of high-resolution spatial data, but their transmission to Earth is limited by severe bandwidth and energy constraints. GRAIL addresses this challenge by developing an end-to-end AI pipeline capable of real-time particle classification—such as electron vs proton discrimination—directly on low-power embedded systems.Speaker: Alberto Garinei (UniMarconi) -
53
GRAIS
The GRAIS project (Gamma-Ray Artificial Intelligence System) aims to develop innovative Deep Learning and Artificial Intelligence (AI) techniques to identify high-energy transient events, such as orphan gamma-ray burst (GRB) afterglows, in data collected by the Fermi Gamma-ray Large Area Space Telescope (GLAST). By leveraging state-of-the-art anomaly detection algorithms tailored to our dataset, we seek to build a model capable of detecting GRBs with atypical characteristics, such as short emission durations or undetected prompt gamma-ray emission. Due to the limited number of observed GRBs, we are also implementing a Generative AI framework to augment the training data by synthesising realistic GRB-like single-photon events, thereby enhancing the performance and robustness of the anomaly detection model.
To date, we have completed a comprehensive exploratory data analysis of both single-photon events detected by the Large Area Telescope (LAT) and catalogued GRB events. We have also defined the architecture for the generative model and developed the clustering algorithm that underpins the anomaly detection model. Ongoing work is focused on finalising both the anomaly detection and generative models, and on fine-tuning the custom evaluation metrics used in the clustering algorithm. These methods have the potential to significantly enhance the discovery of GRBs with diverse properties, offering a novel Deep Learning-driven approach for identifying previously overlooked transient events.
Speaker: Stefano Calì (KOEXAI) -
54
AI-Legs Status Report
Gamma-ray astronomy in the MeV energy range offers the opportunity to investigate important scientific topics, such as nuclear processes, compact objects, cosmic rays and gamma-ray bursts (GRBs). However, in this energy range there is a significant lack of sensitivity compared to other gamma-ray bands. Imaging techniques represent a fundamental tool for signal reconstruction and analysis in this context. The project aims to develop a new image reconstruction algorithm based on advanced machine learning techniques, aimed at interpreting the energy deposits of photons incident on calorimetric modules installed on board space missions. The goal is the precise localization of galactic sources emitting low-energy photons. To achieve this goal, a Monte Carlo simulation of the experimental apparatus was implemented using the Geant4 platform, followed by the development of machine learning models based on CNN and GNN architectures. The presentation will illustrate the current status of the project, the most recent results obtained through Monte Carlo simulation and some proposals under study for classification through Machine Learning techniques.
Speaker: Francesco Conventi (INFN) -
55
Astrodata Status Report
In this update, we would like to provide you with an overview of the list of features that will make up the Astrodata tool, which we have now outlined with precision and detail. The various features will enable the creation of a framework that can be used to robustly apply machine learning algorithms within the workflow of astrophysical research. We will also report on the development progress of each feature.
Speaker: Davide Giannuzzi (Alkemy) -
56
AstroTool: Rearchitecting VisIVO through HPC and Remote Visualization
AstroTool aims to evolve VisIVO into a next-generation client/server platform for the exploration and visualization of large FITS files from SKA and similar facilities. Building on the legacy of the ViaLactea Visual Analytics tool, AstroTool leverages modern High Performance Computing (HPC) infrastructures and a distributed, service-oriented cluster architecture to address the limitations of traditional desktop-based solutions by using a reengineered pipeline for the remote processing of massive FITS datasets.
Speaker: Marco Di Francesco (Alkemy) -
57
Machine Learning and Hardware Acceleration for Space-Based Detection: Particle Track Reconstruction and Low-Energy Gamma Imaging
The Spartan and LEGIMaC projects share a common goal: advancing space-based detection technologies by integrating cutting-edge machine learning algorithms with energy-efficient hardware platforms, such as FPGAs and GPUs. Both projects address critical challenges in the reconstruction and classification of complex physical events in high-rate, noisy environments—typical of astrophysics experiments and orbital detectors.
Spartan focuses on the real-time reconstruction of particle tracks in high-density environments by training deep neural networks on both synthetic and experimental data. The approach leverages temporal information to disentangle overlapping events, enhancing accuracy in track identification. The algorithms are optimized for low-power, radiation-tollerant FPGAs, enabling their deployment in space with minimal energy consumption and high reliability.
On the other hand, LEGIMaC targets the detection of very low-energy gamma events at high rate in space calorimeters, overcoming traditional trigger limitations and the intrinsic dark noise of large-area SiPMs. By applying ML techniques to waveform shape analysis, LEGIMaC discriminates between dark counts and genuine scintillation events with longer decay constants. It also addresses the problem of event pileup in high-flux environments, allowing accurate energy reconstruction at the edge of detectability.
Both initiatives emphasize a hardware/software co-design philosophy to maximize computational efficiency, scalability, and robustness. The potential impact extends far beyond space science, with promising applications in nuclear medicine, security systems, and high-energy physics. Together, Spartan and LEGIMaC represent a major step toward the development of a new generation of intelligent, space-ready detection systems capable of operating in the most challenging conditions.Speaker: Andrea Abba (Nuclear Instruments) -
58
DeepCosmoNet
Understanding the large-scale structure of the universe, known as the Cosmic Web, relies heavily on the analysis of numerical 3D N-body simulations.
Analyzing the output of those simulations remains computationally intensive due to large 3D point clouds and complex clustering tasks.
Traditional methods—such as FoF, SUBFIND, and ZOBOV often operate as separate, CPU-bound tools with limited parallel scalability and require costly post-processing.We introduce a unified DeepCosmoNet framework that segments both high-density (halos and subhalos) and low-density (voids) structures in one GPU-optimized pass. For halos and subhalos, we apply particle-level Graph Neural Networks accelerated by tree‐based nearest‐neighbor searches, avoiding the inefficiencies of voxelization for sparse objects. For voids, we voxelize the density field and employ 3D CNNs framing void identification as a bounding‐box detection problem, with a parallelized 3D Non‐Maximum Suppression.
Our pipeline achieves end-to-end speedups of two orders of magnitude over CPU benchmarks, leveraging hybrid parallelism to maximize throughput.
The approach exhibits near-linear scaling on multi-GPU clusters, demonstrating exceptional throughput and scalability for large simulation volumes.Furthermore, the quality of the detections is notably high; we find that the detections produced by our method are cleaner than the ground truth labels themselves.
By shifting the computational bottleneck from data processing to scientific interpretation, DeepCosmoNet enables accelerated and scalable exploration of the Cosmic Web.Speaker: Vincenzo Del Zoppo (KOEXAI)
-
52
-
Coffee Break
-
Bandi a CascataConvener: Ugo Becciani (Istituto Nazionale di Astrofisica (INAF))
-
59
NeuroStarMap Status Report
In this update, we would like to report on the progress of the NeuroStarMap project, which aims to enable astronomers to obtain more reliable estimates of stellar and cosmological distances. Specifically, we will discuss the neural networks implemented as an initial approach to the project, their results, and the data preparation process that led to them.
Speaker: Stefano Tortora (Alten) -
60
Astrovisio Status Report
In this update, we would like to show the progress of the Astrovisio project regarding the implementation of the API for data management and the development of the Unity application for VR visualization. In particular, we will present the first version of the desktop application and a VR visualization based on FITS data selected in collaboration with a team of astrophysicists from the Scuola Normale Superiore in Pisa.
Speaker: Davide Posillipo (Alkemy) -
61
CANDELA - standard Candle-based Distance Estimation with Learning Algorithm
CANDELA aims to develop a generalized methodology using machine learning and deep learning techniques to estimate the distances of stars and galaxies, leveraging distance indicators available in catalogs such as Gaia DR3 and the OGLE catalogs of variable stars, in combination with parameters extracted from photometric time series. At the same time, by making use of the rich collection of photographic plates from the archive of the Torino Observatory — containing astronomical observations carried out between 1925 and 1996 — another goal is to identify objects of interest within these plates and integrate any relevant objects into the input dataset for the distance estimation models. In this update, we present the progress of this work compared to what was shown at the second technical meeting in Bologna and provide an outlook on the future developments of the project.
Speakers: Andrea Lessio (ITHACA S.r.l.), Virginia Ajani -
62
CUDA-powered Large Scale Investigation of Cosmic Rays Spatial Propagation with Monte Carlo SDEs
In the context of the SDEGnO project, we present recent advancements in the GPU optimization of a Monte Carlo code for spatial propagation.
By implementing modern C++ standard and CUDA libraries, and restructuring the code to evaluate multiple heliosphere parametrizations in parallel,
we achieved an extremely significant speed-up, greatly enhancing the performance of the simulation and its capabilities for parameter exploration.
This talk will cover the optimization methodology, performance benchmarks, and validation results.
Furthermore, we outline the upcoming integration of GPU-optimized reweighting techniques for parameter tuning that will further improve exploration efficiency.Speakers: Leone Bacciu (Università di Venezia), Matteo Grazioso (Università di Venezia) -
63
Clustering of FLUX and EW Maps Across Energy Bands in Supernovae Using Deep Learning Methodologies
Analyzing spectral and spatial information across di.erent energy bands in supernova remnants is crucial for understanding their physical and chemical evolution. In this work, we proposed a novel deep learning methodology aimed at clustering FLUX and Equivalent Width (EW) maps corresponding to di.erent energy bands of individual supernovae. Our approach consists of three main phases: (1) synthetic image generation, (2) embedding computation, and (3) hierarchical clustering. Given the limited number of real images available, we employed Variational Autoencoders (VAEs) [1] to generate synthetic FLUX and EW maps, training dedicated models for each energy band. Subsequently, we computed image embeddings using a Siamese Neural Network [2] with a ResNet18 backbone, optimized to bring embeddings of similar spectral characteristics closer together in the latent space. Embeddings provide a compact and informative representation of the spectral maps, significantly reducing dimensionality and accelerating the clustering process without losing relevant morphological or spectral information. Finally, we performed hierarchical clustering using the cosine distance between embeddings. Comparative experiments against classical clustering approaches based on image linearization with cosine distance and statistically validated clustering [3] with inner product showed that our method achieves clustering results consistent with these traditional techniques. Furthermore, we validated the robustness of the clustering results by applying the same methodology to synthetic data generated from magnetohydrodynamic simulations of supernova remnants [4], confirming the approach's reliability and physical interpretability. This study demonstrates the potential of combining generative and metric learning models to automatically group energy bands according to spectral and morphological similarities. It o.ers a powerful and e.icient tool for the astrophysical analysis of supernova remnants even in low-data regimes. We will also exploit the capability of this approach in discriminating signal from noise in images relative to cosmic background emissions.
Speaker: Prof. Salvatore Micciche (Università degli Studi di Palermo, Dipartimento di Fisica e Chimica Emilio Segrè) -
64
Astroclass Project
The goal of the AstroClass project is to automate the extraction of some characteristic features of astrophysical structures through advanced machine learning techniques, with a focus on extracting density, pressure, and temperature profiles of galaxy clusters from multi-frequency observational data and brightness surface profiles. Leveraging Interpo.latory AutoEncoder neural networks, the project addresses the challenge of extracting meaningful 3D physical profiles from a limited dataset of X-ray and Planck observations.
This talk will present the project’s methodological framework, including data prepa.ration, algorithm design. We will discuss the specific issues encountered due to dataset size and normalization, and the strategies adopted to mitigate overfitting and improve prediction accuracy.
The current validation phase, involving systematic hyperparameter tuning and per.formance evaluation, will be illustrated with preliminary results. Finally, we will outline the next steps towards deploying an open-access platform for the scientific community, designed to facilitate data sharing and collaborative research in the field of galaxy cluster analysis.Speaker: Cristina Martellini (FUKO) -
65
ESPAI
The ESPAI project (Enhancing Signal Purity with Artificial Intelligence in X-band telescopes) aims to develop innovative Deep Learning and Artificial Intelligence (AI) techniques to mitigate contamination from solar flares in X-band astronomical observations conducted by the XMM-Newton telescope. By leveraging state-of-the-art anomaly detection algorithms tailored to our dataset, we seek to build a model capable of distinguishing genuine astrophysical signals from solar flare artefacts. In parallel, we are implementing a Generative AI framework to augment the training data by synthesising realistic solar flare events, thereby enhancing the performance and robustness of the anomaly detection model.
To date, we have completed a comprehensive exploratory data analysis, established the architecture for the generative model, and developed an initial benchmark classifier capable of identifying solar flare photons in a test dataset. Ongoing work focuses on developing a more robust anomaly detection model and generating a synthetic dataset. These methods have the potential to significantly improve signal retention in X-band astronomical observations by enabling the recovery of valuable data that would otherwise be discarded due to solar flare contamination.
Speaker: Stefano Calì (KOEXAI)
-
59
-
13:15
Lunch
-
