IN2P3/IRFU Machine Learning workshop 2026

16 nov. 2026, 10:00 → 18 nov. 2026, 16:00 Europe/Paris

Amphitheatre (LPSC)

General Information

We are pleased to announce the next edition of our annual IN2P3 & CEA-IRFU AI for Physics Workshop.

The workshop will take place in person at LPSC Grenoble (details on how to reach the LPSC here) from Monday 16 November (11:00) to Wednesday 18 November (15:00). The workshop will also be broadcasted live on Zoom.

The workshop covers all aspects of Artificial Intelligence applied to accelerator science, astroparticle physics, cosmology, detector instrumentation, nuclear physics, particle physics, scientific computing, theoretical physics, etc. It aims to foster exchanges between researchers developing or applying AI methods across these domains. On-going ML activities at IN2P3 are updated on https://machine-learning.in2p3.fr.

The call for abstracts is now open and will remain open until 20 September 2026.

We welcome contributions presenting:

• recent scientific results,
• work in progress,
• new ideas and methods,
• software and infrastructure,
• benchmark datasets and community initiatives.

 

As in previous editions, we particularly encourage presentations of ongoing work and discussions of future directions, in addition to published results.

Note: Results from large collaborations may require approval from the relevant collaboration board. Since this is a national workshop, presentations by students and early-career researchers on ongoing work may be possible; please consult your collaboration's national contact if approval is required.

Tracks

1. AI for Physics Analysis, Inference and Discovery
2. Foundation Models, Representation Learning and Generative AI
3. Trustworthy AI for Physics
4. Signal Processing, Calibration and Reconstruction
5. Event/Object Reconstruction and Identification
6. Real-Time AI, Trigger and Edge Computing
7. AI for Scientific Instruments and Accelerators
8. AI Ecosystem: Infrastructure, Open Science and Community

Mailing list

Please make sure you are subscribed to MACHINE-LEARNING-L@in2p3.fr on IN2P3 listserv and visit our website machine-learning.in2p3.fr to keep up to date with ML at IN2P3. 

Organisation

Organisers

• Valérie Gautard (CEA/IRFU)
• David Rousseau (IJCLab/IN2P3)
• Thomas Vuillaume (LAPP/IN2P3)

Local organising committee (LPSC): 

Johan Bregeon, Sabine Crépé-Renaudin, Pierre-Antoine Delsart, Cyrille Doux, Fabian Lambert, Juan Macias-Perez, David Maurin, Laurent Neiger, Jérome Odier, Carlo Pandini
To reach the LOC, please write to mlworkshop2026-loc@lpsc.in2p3.fr

Support

We acknowledge the support of LPSC,  UGA and MIAI:

lun. 16 novembre

11:30 → 13:30 Registration and welcome lunch

13:30 → 14:10 Welcome address: Accueil du LPSC

13:30 Welcome address by the LPSC director

Orateur: Laurent Derome (LPSC)

13:50 Welcome address by local organization committee

14:10 → 15:10 Signal Processing, Calibration and Reconstruction

AI methods for detector-level signal processing and detector response modeling.

Examples include:
- Signal reconstruction and denoising
- Waveform analysis and deconvolution
- Detector calibration and alignment
- Differentiable calibration
- Detector response correction

14:10 Multi-Parameter Calibration and Uncertainty Propagation in LArTPCs Using Differentiable Simulation

Precision measurements in Liquid Argon Time Projection Chambers (LArTPCs) rely on accurate modeling of detector response. In standard calibration workflows, effects such as space charge, recombination, electron lifetime, and electronics response are usually measured independently. This segmented approach ignores correlations among parameters and can introduce systematic biases into the simulation.

Differentiable simulation offers a direct way to address this limitation by providing gradients through the full detector model, enabling all parameters to be optimized at once. Using a LArTPC simulator implemented in JAX, we demonstrate simultaneous multi-parameter calibration on realistic, data-like samples. We discuss practical challenges encountered when moving beyond toy setups, including parameter degeneracies, sample statistics, and detector noise.

In addition, we show how automatic differentiation simplifies systematic uncertainty propagation. Because the simulator provides exact Jacobians and Hessians, the full parameter covariance matrix can be mapped directly to reconstructed physics quantities without relying on computationally expensive re-simulations.

Orateur: Pierre Granger (CERN)

14:30 Deep Learning for Extracting Spectra form JWST Spectrograms

With telescopes like the Euclid in orbit and CSST in the works, we are looking at an emerging era of astronomy with surveys producing millions and millions of slitless spectrograms. While the analytical extraction techniques are sufficient for the current amount of data, the community will require faster pipelines for future data releases. Added to the large amount of expected data, slitless spectroscopy, as executed by instruments such as Euclid/NISP, presents significant challenges for data processing — most notably spectral contamination arising from the overlap of multiple dispersed sources on the detector.
To address these difficulties, we are developing a machine learning approach aimed at extracting and decontaminating 1D spectra directly from real flight 2D spectrograms. Our strategy proceeds incrementally; as a first step, we develop and validate a pipeline to extract 1D spectra from 2D JWST slit spectroscopic data. In the subsequent stages, we will match the spectrogram resolution and wavelength range to that of Euclid, we will emulate the spectral overlap artificially to then train a model to undo said overlap, we will automate the identification of first orders in the NISP spectroscopic images, and finally test the end-to-end pipeline on real Euclid data.
In the first stage, we train a 3-branch attention-inception CNN that can estimate 1D spectra from 2D spectrograms with no information about observed wavelengths or resolution. Each branch estimates different parts of the spectrum, like the continuum, the continuum-free component and refinements. To achieve this performance, we divide spectrograms into smaller segments, allowing the network to learn the 2D-1D mapping with no regard for object properties, wavelength or resolution.
While the high-level goal is tailored for Euclid data, the strategies used in this work will serve as a blueprint for any future slitless or even slit spectroscopic surveys. This generality makes this work a valuable asset in the coming era of large-scale, slitless and highly multiplexed spectroscopic surveys, where efficient and robust automated extraction pipelines will be essential.

Orateur: Soorya Narayan Rajeshkumar (Aix Marseille Univ, CNRS/IN2P3, CPPM, Marseille, France)

14:50 A plug-and-play approach for tomographic weak-lensing mass mapping

Upcoming surveys such as Euclid and Rubin will probe the matter distribution through weak lensing with unprecedented precision. However, reconstructing mass maps from noisy shear measurements is fundamentally an ill-posed inverse problem. While current deep-learning approaches achieve high reconstruction accuracy, they either must require retraining for each new sky region (due to varying mask and noise setting), or rely on slow MCMC sampling.

We introduce PnPMass, a plug-and-play method that alternates between a gradient step on a noise-whitened data-fidelity term, and a denoising step performed by a single neural network trained on simulations corrupted by Gaussian white noise. On simulated mass maps, it achieves accuracy close to the best deep-learning methods, converges in a few iterations, and requires no retraining when the mask and noise setting change. We then extend the method to tomographic mass mapping in order to resolve structures along the line of sight. Because the lensing signal is strongly correlated across redshift bins, we propose to apply a joint reconstruction by replacing the denoiser with a multichannel network. In preliminary tests on simulations, this approach, referred to as PnPMass3D, reduces the root-mean-square error by 20% over applying PnPMass to each bin independently. Furthermore, the gap is much larger once the mass estimates are disentangled across redshift bins with the Bernardeau-Nishimichi-Taruya (BNT) transform: independent reconstruction shows spurious foreground contamination in higher-redshift bins, whereas PnPMass3D recovers much more accurate per-bin mass estimates.

Link to the published paper: https://www.aanda.org/articles/aa/abs/2026/06/aa57652-25/aa57652-25.html. Note: the tomographic extension is not included; this is a work in progress.

Orateur: Hubert Leterme (CEA IRFU/DAp)

15:10 → 15:30 Real-Time AI, Trigger and Edge Computing

Machine learning methods operating under latency, bandwidth, and hardware constraints.

Examples include:
- Trigger systems
- FPGA and embedded AI
- Online event selection
- Fast inference
- AI at the edge
- Resource-aware machine learning

15:10 R&T Think2

We will present the status of the IN2P3 R&T THINK2, whose objective was to develop tools to embed machine learning algorithms in FPGAs for realtime applications (DAQ, trigger,...)

Orateur: Sebastien Viret (IPN Lyon)

15:30 → 16:00 ☕

16:00 → 17:25 AI for Scientific Instruments and Accelerators

AI methods for the design, optimization, operation, and control of scientific facilities.

Examples include:
- Accelerator diagnostics and control
- Detector and instrument optimization
- Surrogate models
- Digital twins
- Experimental design
- Scientific control systems

16:00 Physical and ML models for simulating spectroscopy instruments

In the ANR DISPERS project, we use a physics-first approach to simulate spectroscopy instruments. Each physical stage of the instrument, such as optics, source catalogs, detector response, and calibration, is modeled based on physical principles wherever possible. Machine learning is only used to complement these models where physics alone is insufficient or too expensive. These model components are integrated through Simcraft, a modular simulation framework built around interchangeable, pluggable physical model components. We present three such components, each of which addresses one stage of the simulation.

• Synthetic catalog generation samples from AION, a multimodal transformer foundation model (Parker, Lanusse, Shen et al., arXiv:2510.17960, 2025), replacing hand-tuned parametric distributions with catalogs that inherit realistic cross-modal correlations (fluxes, morphology, spectra). Grew from foundation-model benchmarking at the AstroInfo 2025 hackathon into a package via a CentraleDigitalLab student project.

• PSF modeling pairs a physics-based forward simulator (aperture geometry, Zernike aberrations, polychromatic sources) with ML models inverting the problem: Zernike/wavefront-error coefficients estimated directly from a PSF image, without a dedicated wavefront sensor. We combine a discriminative CNN regressor (Sauniere, Gillard & Zoubian, SPIE 2024) with a DDPM diffusion model that also reconstructs the PSF with calibrated uncertainty (SPREAD; Sauniere et al., in prep. for AAS/AJ).

• Instrument calibration couples a geometrical-optics dispersion model with an ML residual corrector as an interpretable Chebyshev-polynomial alternative. Benchmarked against a published on-ground result (Gillard et al., 2025, arXiv:2506.08378): preliminary results are the same order of magnitude but do not yet match its accuracy.

All three are validated against Euclid/NISP ground-test data as a case study, but the approach, physical model plus targeted ML correction, and the simulation framework, generalizes beyond this instrument.

Orateur: Julien Zoubian (Aix Marseille Univ, CNRS/IN2P3, CPPM, Marseille, France)

16:20 Unsupervised Machine Learning for Anomaly Detection in Beam Position Monitors Data

Turn-by-turn beam position monitors (TbT-BPMs) are the primary
diagnostic for optics measurement and correction in high-luminosity colliders.
Reliable identification of faulty sensors is essential to any optics
reconstruction pipeline, yet manual inspection does not scale to the hundreds
of monitors in current rings or the thousands expected in future machines such
as the FCC. We apply two complementary unsupervised machine-learning methods
to TbT-BPM data from both rings of SuperKEKB: Density-Based Spatial Clustering
of Applications with Noise (DBSCAN) with automatic feature selection via the
Time2Feat library, and a one-dimensional convolutional autoencoder (1D-CNN AE)
that operates directly on raw data without labelled training
examples. Both methods are evaluated on 2024 data against a
reference list of hardware-damaged BPMs provided by the SuperKEKB operations
team. On the High Energy Ring (HER), DBSCAN detects faulty BPMs with a recall (TPR) of 83\%, while the 1D-CNN AE achieves 80\%. On the Low Energy Ring (LER), the AE recovers all known-faulty BPMs (100\% TPR) while
DBSCAN recovers three (50\%). The models further identify strong candidates for unreported hardware faults, ranking them among
the top outliers for the HER and LER measurement sessions.

Orateur: Quentin Bruant (CEA/IRFU/DACM)

16:40 ARTIFACT network

The ARTificial Intelligence For Accelerators, user Communities and associated Technologies (ARTIFACT) network—a collaborative, open-science network connecting research infrastructures, institutes, and universities— addresses these challenges by coordinating a concerted strategy for AI in accelerator science. Beyond accelerator physics, ARTIFACT establishes bridges with adjacent disciplines, including medical applications and core AI research on interpretability and trustworthiness. This contribution presents the network’s vision, reference architecture, and strategic pillars organized around data and models, sustainability, current and future infrastructure operation, and training. It also outlines a roadmap for 2026–2031 that addresses key scientific, technical, and organizational challenges for sustainable AI adoption in Particle Accelerators research and technological infrastructures.

Orateur: Dr Adnan Ghribi ({CNRS}UPR3266)

17:00 The iRIS project : intelligent research infrastructures

iRIS positions Research Infrastructures as testbeds and innovation engines for the green and digital transitions . Motivated by the Future Circular Collider and the Einstein Telescope, this project aims at demonstrating AI-powered environmental technologies and sustainability-by-design across the infrastructure lifecycle. Its relevance extends beyond fundamental-science RIs: AI approaches for particle accelerator design and optimisation can improve resource efficiency in industrial plants. AI-powered real-time materials characterisation and separation can accelerate circular construction and recycling. The soil reconstitution pilots advance land restoration and ecosystem service recovery. These technological demonstrations are integrated with a harmonised socio-economic sustainability appraisal combining lifecycle and cost-benefit approaches to support evidence-based investment decisions . Outreach, training and open tools strengthen European skills, industrial uptake and societal trust. By federating the European Commission’s Joint Research Centre and integrating with the RAISE efforts towards AI driven infrastructures for science, iRIS ensures that European sustainability concepts, indicators and policy frameworks are integrated from the outset, enhancing replicability, relevance of policy feedback to the EC, and establishing a European reference model for sustainable design and operation of research infrastructures.

Orateur: Dr Adnan Ghribi ({CNRS}UPR3266)

17:20 Machine learning surrogate models for collective effects studies in the FCC-ee High-Energy Booster

Accurate beam dynamics simulations including collective effects are essential for the design of future particle accelerators, but high-fidelity macroparticle tracking can become computationally expensive, especially when large parameter scans are required. We investigate machine learning surrogate models for the evolution of beam distributions in the FCC-ee High-Energy Booster.

Rather than predicting the trajectories of individual macroparticles, the approach learns the evolution of the beam distribution itself. Different architectures are explored, including Fourier Neural Operators, conditional variational autoencoders, Transformers, and graph-based neural operators, using data generated from accelerator tracking simulations. Particular attention is given to preserving the physical structure of the accelerator and to learning local interactions along the machine lattice. The results illustrate how structure-aware machine learning can provide fast approximations of beam distribution evolution. This suggests a promising route toward more efficient computational studies of collective effects in beam dynamics.

Orateur: Keon Hee Kim (CNRS/GANIL)

mar. 17 novembre

09:00 → 10:25 AI Ecosystem: Infrastructure, Open Science and Community

Tools, infrastructures, datasets, and community initiatives enabling AI research in physics.

Examples include:
- Research software
- AI frameworks and workflows
- HPC and distributed computing
- FAIR data and open science
- Benchmark datasets and challenges
- Community initiatives and training

09:00 Jean Zay for AI in Physics: Resources, Access and Getting Started

Jean Zay, hosted at IDRIS, provides large-scale computing resources for AI research. This talk will offer a practical introduction for the IN2P3 and CEA-IRFU communities, focusing on the resources available and the steps required to access and use them effectively.

I will present the computing resources and software environment, together with an outlook on future capabilities. The presentation will address both administrative and technical aspects of access: eligibility, allocation schemes, proposal preparation and evaluation, and the practical steps from obtaining an account to running AI workloads. Particular attention will be paid to demand for computing time and its implications for applicants, including how to estimate computational needs and formulate realistic, well-justified resource requests. I will also discuss practical considerations for GPU-based and distributed workloads, as well as the user support and training available at IDRIS.

A few examples drawn from cosmology, including emulation and simulation-based inference, will illustrate relevant use cases and their computational requirements. The aim is to help participants assess whether Jean Zay is a suitable platform for their AI projects and identify a clear path from an initial idea to a computing allocation and their first runs.

Orateur: Stéphane Ilic (IDRIS)

09:30 LLM and agents

Orateurs: François Portet (IMAG), Maxime Calka (UGA)

10:00 Lightcone Research: an open ecosystem for reproducible, composable and verifiable physics analyses in the age of AI agents

Nowadays, AI agents that can write and run scientific code on their own. From coding assistants such as Claude Code, Cursor and Codex to integrated research systems such as Kosmos and Gemini Co-Scientist or Claude Science, they all are rapidly changing how analyses are built in physics. This speed also makes an existing problem worse. Methodological choices, assumptions and scientific justifications stay buried in code that is hard to inspect, reproduce or extend. For communities such as particle physics, astroparticle physics and cosmology, results depend on long chains of analysis decisions, and without suitable infrastructure AI-assisted research risks deepening the reproducibility crisis rather than solving it.
We present Lightcone Research, an open-source initiative, anchored between UC Berkeley and the CNRS, that aims to build the missing ecosystem for reproducible, composable and verifiable research in this new setting. At its core is ASTRA, an open declarative format that captures the scientific intent of an analysis in a standard that both humans and AI agents can read. That intent covers the problem, the inputs, the decisions, the analysis parameters, the outputs. Analyses that follow ASTRA are reproducible by default. More importantly, they are inspectable, composable and citable. Every result can be traced back to the assumptions that produced it, and individual decisions can be swapped without rewriting the whole pipeline. Analyses can also reference and build on one another.
The ecosystem is organised in four layers: (1) the ASTRA protocol for reproducible science; (2) an agentic layer for assisted analysis authoring, that works with (3) a command-line interface for validating analyses and running them seamlessly on HPC infrastructure, including NERSC and French computing centres; and (4) a modern web platform based on the Jupyter ecosystem for inspecting the results and their full provenance. We describe the architecture and showcase the platform at CC-IN2P3 and the reproductions of recent analyses in cosmology. We also report lessons learned from a beta-testing phase with researchers and discuss how the adoption of such system could be spread across IN2P3 and IRFU communities.

Orateur: Alexandre Boucaud (APC / IN2P3)

10:20 French tokens for everyone : partnering with the ILaaS Federation

Large language models are quickly becoming everyday tools for physicists and software developers, whether for coding help, documentation, data exploration or agentic workflows. Yet access is uneven, and most usage goes through shadow IT with commercial frontier APIs, which raises concerns about cost, data sovereignty and environmental footprint. Many research needs, however, do not require a frontier model: well-chosen open-weight models, served efficiently on shared infrastructure, cover a large share of daily use cases.
This lightning talk introduces the ILaaS Federation and its partnership with CC-IN2P3, which aims to make frugal, sovereign AI inference available to the whole research community. We explain how users obtain their API keys, which models are currently available, and how per-user compute tracking gives visibility on actual resource consumption, supporting both fair sharing and a more sober use of AI. We close with concrete use cases and a call for new partners and users.

Orateur: Alexandre Boucaud (APC / IN2P3)

10:25 → 10:55 ☕

10:55 → 12:35 AI for Physics Analysis, Inference and Discovery: Part 1

Machine learning methods supporting scientific measurements, interpretation, and discoveries.

Examples include:
- Statistical inference
- Parameter estimation
- New physics searches
- Scientific time-series analysis
- Event classification
- Scientific interpretation

Président de session: Dr Thomas Vuillaume (LAPP, Univ. Savoie Mont-Blanc, CNRS)

10:55 Accelerating Weather Forecasts with Artificial Intelligence

Until very recently, weather forecasting almost only relied on solving physical equations using numerical solvers called numerical weather prediction (NWP) models. Modelling the atmosphere faster and at higher precision was made possible thanks to access to more computing power through the evolution of High-Performance Computing (HPC) clusters.
The rise of Artificial Intelligence (AI), more specifically neural networks, changed the game. 2022 marked an important milestone with the development of AI models emulating the whole atmosphere, thanks to the huge amount of simulation data available. These data-driven models have improved constantly since, showing very competitive scores against the state-of-the-art of operational global models, consuming ten to hundred times less energy during their lifecycle. Developing these AI models alongside NWP ones is thus crucial for weather forecast centers on the road to faster and more accurate weather predictions.

Orateur: Dr Mikaël Jacquemont (Bull SAS)

11:25 emulating accurately and consistently the cosmological power spectrum and the halo mass function

Cosmological forecasts and likelihood analyses increasingly need two ingredients at once: the linear matter power spectrum, and the abundance of dark matter haloes. Computing them exactly is far too slow for parameter inference, and computing them from unrelated approximations quietly introduces inconsistencies between the two.

In this presentation, I will introduce two emulators built to solve both problems together: emu_pk and emu_hmf, in their first public releases. Both are fast, both are automatically differentiable — so gradients with respect to cosmology come for free — and both are designed to be used side by side.

I will show the accuracy they achieve on cosmologies never seen during training. emu_pk reproduces the power spectrum to about 0.1 %, over a wider range of cosmologies than comparable public emulators, including massive neutrinos and evolving dark energy, and it recovers the derivatives that forecasts actually consume — two of them exactly. emu_hmf recalibrates the standard Tinker (2008) halo mass function as a function of cosmology and redshift, reducing its error from roughly 7 % to 0.5 %, a factor of more than ten.

I will close on why consistency matters as much as accuracy: the two emulators share the same conventions and the same definition of the matter field, so going from cosmology to a halo mass function is a single, coherent, differentiable chain. Both are publicly released and installable in one line.

Orateur: johan comparat

11:45 Speeding-up the search for anomalies in ZTF time series with interpretable machine learning

Machine learning is often viewed as a black box when it comes to understanding its output, be it a decision or a score. Automatic anomaly detection is no exception to this rule, and quite often the astronomer is left to independently analyze the data in order to understand why a given event is tagged as an anomaly. Interpretable AI on the other hand provides clues to the analyst, often in the form of feature importance. However, generalist methods on the market tends to exhibit slow performances when applied to big data.

Here I’ll present how the Signatures method, a post-hoc method-specific scheme providing a surrogate anomaly model, can reach tractable performances when applied on the 720 million light-curves from ZTF DR23 and enable active learning for feature selection and unknown unknows searc, therefore improving the rate at which anomalies are found in the sample.

Orateur: Emmanuel Gangler (LPC)

12:05 AI/ML for large galaxy survey analyses

Large galaxy surveys such as the Dark Energy Survey (DES) or the new LSST from the Vera C. Rubin Observatory offer several cosmological probes to test dark energy and gravity. The combination of galaxy clustering and weak lensing has proven to be a powerful probe, especially in combination with complementary probes such as galaxy clusters, supernovae and the baryonic acoustic oscillation peak. After describing the weak lensing cosmological analysis pipeline we validated for the full DES (arXiv:2601.14859), I will highlight the main challenges, from computing costs to the control of parameter biases, and the latest developments to tackle them with ML and LLMs, including emulators, samplers and automated benchmarks. I will end with an application of LLMs to Vera C. Rubin Observatory data visualization.

Orateur: Agnès Ferté (LPNHE)

12:25 Revisiting Nuclear Cross Sections for Cosmic Rays with Gaussian Processes

Inelastic and production cross sections are crucial inputs for predicting nuclear cosmic-ray fluxes in the Galaxy. Deuterium (²H) and helium-3 (³He) are both expected to be produced solely by interactions of heavier nuclei with the interstellar gas, yet recent AMS-02 and Voyager data are in tension with this picture. In this context, it is natural to revisit the relevant nuclear cross sections. Using several compiled datasets (EXFOR, HEPData, PDG), we derive new data-driven nuclear models based on 1D or 2D Gaussian processes. This method provides uncertainties together with their correlations.

Orateur: Anthony Correia (LPSC,IN2P3/CNRS/UGA)

12:30 Physics-informed impedance-model-agnostic reconstruction of point-charge wakefields from finite bunch electromagnetic simulations

Electromagnetic wakefield simulations generally provide wake potentials excited by finite length bunches, whereas beam dynamics studies require the underlying point-charge wake and broadband impedance. Direct access to the THz impedance regime in full 3D electromagnetic simulations typically requires extremely short drive bunches and very fine numerical resolution. The resulting simulations can be computationally expensive and, at sufficiently short bunch lengths, may also become numerically fragile in ways that are not always resolved by further mesh refinement. This motivates reconstructing the point-charge response from simulations performed with longer, computationally more tractable bunches.

We develop a impedance-model-agnostic reconstruction method, PCWR, which infers the point-charge wake directly from finite bunch electromagnetic simulations data without assuming an analytical impedance model. The method combines Golub–Kahan Krylov inversion, regularized solution ensembles, uncertainty based spectral trust estimation, and physics-informed reconstruction of poorly constrained frequency regions. Causality and consistency with the simulated finite bunch wake are enforced throughout the reconstruction. The method is tested on three qualitatively different structures: a single-layer resistive wall chamber, and a two-layer NEG/Cu resistive wall chamber, and a corrugated dechirper. The PCWR provides a route to extracting short-range wake and high frequency impedance information that is inaccessible or unstable under straightforward deconvolution.

Orateur: Keon Hee Kim (CNRS/GANIL)

12:35 → 14:00 Lunch

14:00 → 14:50 AI Ecosystem: Infrastructure, Open Science and Community

Tools, infrastructures, datasets, and community initiatives enabling AI research in physics.

Examples include:
- Research software
- AI frameworks and workflows
- HPC and distributed computing
- FAIR data and open science
- Benchmark datasets and challenges
- Community initiatives and training

14:00 Beyond efficiency: Frugal AI

The adoption of AI is accelerating and its ecological impacts are soaring, despite tremendous improvements of technologies efficiency. Indeed, such improvements have historically consistently led to increased usage and thus overall impacts of AI technologies. Yet, frugality of AI is a major preoccupation, e.g. at the national (AFNOR reference document), and European (AI act) levels. In this context, how to include frugality in an AI project ? As a first step, practical approaches exist to assess the various impacts of a AI project, from data collection to model use. Then, methodologies can be adopted to effectively reduce those impacts: looking for alternatives, good practices. While frugal AI is a necessary pathway, it still shows limits regarding systemic effects.

Orateur: Dr Sophie CERF (INRIA)

14:30 MIAI Cluster: the Grenoble Alpes AI cluster and what it offers to physics

MIAI Grenoble Alpes is one of the nine French AI clusters, built on the 3IA institute created in 2019 and coordinated by Université Grenoble Alpes with UCA and USMB. It brings together about 600 researchers from more than 100 laboratories, from machine learning, generative AI and explainable or frugal AI to vision, robotics, health, environment and industry. This talk presents the cluster's research axes, its training programme (EFELIA-MIAI), its industrial partnerships and its calls for projects, and discusses how physics and instrumentation communities can connect with it.

Orateur: Cyrille Doux (LPSC)

14:50 → 16:00 AI for Physics Analysis, Inference and Discovery: Part 2

Machine learning methods supporting scientific measurements, interpretation, and discoveries.

Examples include:
- Statistical inference
- Parameter estimation
- New physics searches
- Scientific time-series analysis
- Event classification
- Scientific interpretation

Président de session: Dr Thomas Vuillaume (LAPP, Univ. Savoie Mont-Blanc, CNRS)

14:50 Hybrid Latent-Space Inference for Galaxy Deblending

In the absence of systematic errors, weak lensing is among the most constraining probes of dark energy. Blending is a major source of systematic uncertainty for weak-lensing analyses, affecting both galaxy shape measurements and photometric redshifts. At LSST depth, blending is estimated to affect about 62% of detectable galaxies. Deblending therefore requires assigning the observed flux among the individual sources contributing to a scene. Because this source-separation problem is intrinsically underconstrained, deblending methods differ mainly in the assumptions or priors used to regularize the decomposition. Deep-learning methods instead learn galaxy representations directly from data, reducing the need for explicit parametric assumptions about galaxy morphology. Generative approaches can learn a prior from isolated galaxy images and use it to constrain source separation and the reconstruction of individual galaxy images from blended scenes. This raises a central question: whether image reconstruction is itself the most informative endpoint for recovering galaxy properties, or whether these properties can be inferred more accurately directly from the learned latent representation.

We investigate this question using a two-stage variational autoencoder framework based on DebVader. A VAE is first trained on isolated galaxies to learn a generative latent representation, after which the decoder is frozen and a second encoder is trained to map blended observations into this latent space. Our main contribution is a hybrid training strategy in which ellipticity and photometric redshift are predicted directly from the latent representation through supervised MLP heads while image reconstruction is retained as a joint objective.

We find that ellipticity and photometric redshift can be accurately inferred directly from the latent representation of blended scenes. Direct latent-space ellipticity estimation outperforms measurements on reconstructed images, and increasing the relative weight of the ellipticity and redshift prediction losses further improves parameter recovery. At the highest tested weight on the prediction losses, the hybrid model also improves ellipticity recovery from reconstructed images.

Orateur: Anna Preto

15:10 Accurate and efficient simulation-based inference for massive black-hole binaries with LISA

The Laser Interferometer Space Antenna (LISA) is a future space-based gravitational-wave observatory designed to detect low-frequency gravitational waves, including those produced by massive black-hole binaries. I will present a machine-learning framework based on normalizing flows for rapid Bayesian parameter estimation of these sources. The method is trained on simulated LISA signals and can generate thousands of posterior samples in seconds, while maintaining good agreement with traditional inference methods over a broad range of signal-to-noise ratios. I will discuss the performance and current limitations of the approach, as well as future developments towards more realistic LISA data analysis.

Orateur: Alice Spadaro (L2I Toulouse, CNRS/IN2P3, Université de Toulouse)

15:30 MODULARITY: A Modular Architecture for Distributed Autonomous Scientific Discovery

Modern scientific facilities increasingly combine complex instrumentation, simulations, heterogeneous diagnostics, and machine-learning-based control. We propose MODULARITY, a domain-agnostic architecture for distributed closed-loop experimentation, where complex systems are decomposed into interoperable autonomous modules. Each module combines observations, control variables, constraints, a local digital twin, and quantified uncertainty. Local predictions and measurements are continuously aggregated into Global Maps, dynamic representations of the experimentally accessible space containing current performance, explored and unexplored regions, model uncertainty, physical constraints, and potentially scientific novelty. These maps provide the common information layer through which higher-level agents coordinate and guide local optimization, rather than relying on a single monolithic model or controller \cite{tennant2026autonomous,brynes2026closing,kelly2026bayesian,hickman2025atlas}.

The ML layer combines probabilistic surrogate modelling, Bayesian inference, Bayesian optimization, reinforcement learning, and decentralized or multi-agent optimization. Local agents perform data-efficient control and adaptation, while Global Maps act as an evolving optimization and exploration landscape, identifying where additional measurements or actions are expected to provide the greatest operational or scientific value. The coordination layer can therefore allocate optimization effort across modules according to predicted gain, uncertainty, constraints, and cross-module dependencies. Beyond performance optimization, MODULARITY uses Bayesian experimental design to guide scientific inference, selecting experiments that maximize expected information gain,

$$ d^\star=\arg\max_d I(\theta;Y\mid d,D), $$ or discriminate between competing models through $I(M;Y\mid d,D)$ \cite{rainforth2024bed}. Global Maps can consequently evolve from simple performance landscapes into probabilistic knowledge maps combining model predictions, uncertainty, experimental coverage, and evidence for model inadequacy.

At the discovery level, MODULARITY distinguishes parameter inference, model discrimination, persistent model inadequacy, and open-world novelty. Unexpected observations update the Global Maps and can redirect the autonomous exploration toward regions where competing models diverge, uncertainty is large, or observations systematically disagree with predictions. Such regions trigger targeted replication, diagnostic cross-checks, nuisance-parameter inference, and experiments designed to discriminate between alternative explanations. Scientific value can therefore be represented through a multi-objective discovery utility combining information gain, falsification power, predictive surprise, reproducibility, novelty, cost, and risk. Domain-specific modules define the relevant likelihoods, priors, observables, and evidential criteria, while the common Global Map and coordination framework enables applications ranging from particle accelerators and particle physics to materials science.

References

\cite{rainforth2024bed}
T. Rainforth, A. Foster, D. R. Ivanova, and F. B. Smith, “Modern Bayesian Experimental Design,” Statistical Science 39(1), 100–114 (2024).

\cite{hickman2025atlas}
R. J. Hickman et al., “Atlas: a brain for self-driving laboratories,” Digital Discovery 4, 1006–1029 (2025).

\cite{tennant2026autonomous}
C. Tennant, “Toward a Fully Autonomous, AI-Native Particle Accelerator,” arXiv:2602.17536 (2026).

\cite{brynes2026closing}
A. D. Brynes et al., “Closing the Loop: Deploying Auto-Generating Digital Twins for Particle Accelerators,” arXiv:2604.19101 (2026).

\cite{kelly2026bayesian}
C. Kelly et al., “Machine Learning Assisted Bayesian Calibration of an Accelerator Digital Twin from Orbit Response Data,” arXiv:2608.17820 (2026).

Orateur: Hayg Guler (IJCLAB)

15:50 Machine Learning-Based Event Categorisation for Higgs-to-Dimuon Analyses

The search for the Higgs boson decay to a pair of muons provides a direct probe of the Higgs coupling to second-generation fermions and represents an important test of the Standard Model. Due to the small branching fraction of this decay and the large backgrounds from Drell–Yan and other processes, achieving high sensitivity requires efficient exploitation of the kinematic and event-level information available in the data.
In this talk, we present a machine-learning-based strategy for the Higgs-to-two-muon analysis, employing both Boosted Decision Trees (BDTs) and Neural Network architectures to classify signal and background events. The resulting discriminants are used to categorise events into several mutually exclusive and collectively exhaustive categories, designed to target different production modes and regions of signal-to-background sensitivity. This categorisation enables the analysis to exploit complementary event topologies and kinematic information while maintaining statistically independent event samples.
The analysis is then performed simultaneously across all the different categories, combining their individual contributions to maximise the overall sensitivity to the Higgs-to-two-muon signal. We discuss the role of the machine-learning models in the event categorisation, the optimisation of the category scheme, and the impact of this approach on the sensitivity of the measurement.

Orateur: Somanko Saha

16:00 → 16:30 ☕

16:30 → 17:45 AI for Physics Analysis, Inference and Discovery: Part 3

Machine learning methods supporting scientific measurements, interpretation, and discoveries.

Examples include:
- Statistical inference
- Parameter estimation
- New physics searches
- Scientific time-series analysis
- Event classification
- Scientific interpretation

Président de session: Dr Thomas Vuillaume (LAPP, Univ. Savoie Mont-Blanc, CNRS)

16:30 Self Organised Randoms for correcting systematic effects in Euclid photometric galaxy clustering.

The Euclid satellite is designed to make accurate measurements of weak lensing and galaxy clustering for the purpose of probing structure formation over half of the age of the Universe. However, spatial variations in the Euclid survey and instrumental properties, as well as on the foreground sky emission, can affect photometric galaxy clustering and bias cosmological results. To correct for these systematic effects, we have developed a self-organising map (SOM) algorithm based on the KIDS collaboration's previous work. This algorithm enables us to produce Self-Organised Randoms (SORs) that mimic the systematic contribution. We will present the algorithm's main characteristics and discuss its performance in realistic simulations of Euclid observations.

Orateur: Dr Jérôme Odier (LPSC Grenoble)

16:50 LIS² (Large Image Split Segmentation): A Modular Research Software Framework for Deep Learning on Large-Scale Scientific Images

The automatic detection of interstellar filaments in large-scale astrophysical surveys is a challenging segmentation problem for deep learning methods. Observations cover large portions of the Galactic plane and exhibit a wide range of densities and contrasts, spatially varying noise, and a heterogeneous distribution of structures. In addition, available annotations are necessarily incomplete, making the evaluation of segmentation models over the full observation particularly challenging.

LIS² (Large Image Split Segmentation) was developed to address these constraints through a dedicated workflow for deep learning on very large scientific images. The toolbox handles image splitting and preprocessing, data storage, model training and inference, as well as model evaluation through cross-validation. Particular attention is given to spatial partitioning: training and evaluation data are spatially separated to prevent overlap between regions while preserving the diversity of structures present in the observation. The K-fold framework then provides evaluation over several independent partitions, ensuring a robust comparison of model performances.

LIS² has been used to implement and compare segmentation architectures such as U-Net and its extensions on observations covering large portions of the Galactic plane. The framework also allows physical information to be incorporated into the models: encoding Galactic position in the network makes it possible to account for variations in the context along the Galactic plane and improves the detection of low-contrast structures.

The entire toolbox is modular and configurable, from data processing to training and inference, including models and evaluation procedures. This design makes it possible to adapt the workflow to the characteristics of the data and to the methodological choices of each problem while maintaining a reproducible experimental framework. LIS² provides a Research Software infrastructure for deep learning on large-scale scientific images, combining scalable data processing, unbiased spatial evaluation and configurable AI workflows.

Orateur: Dr Dominique Benielli (Aix Marseille Univ, CNRS, LIS, Marseille, France)

17:10 Cosmology using topological data analysis with SBI

In this talk I will present a cosmological analysis of weak lensing data from the Dark Energy Survey using a topological descriptors, persistent homology, within a simulation-based inference framework. This analysis captures cosmological information from the field beyond Gaussian, complementing standard two-point correlation functions.

Orateur: Cyrille Doux (LPSC)

17:30 Lorentz-Equivariant Transformers for ttX multi-lepton final state at the LHC

Rare top-quark processes, ttW,ttH,tH,ttZ,tttt, are traditionally measured in separate analyses, despite substantial overlap among their multilepton final states. However, this strongly limits the overall physics understanding, as the correlations between the various measurements are large and poorly known. We propose a new approach to tackle these limitations by simultaneously reconstructing and classifying these processes within a unified framework. The reconstruction of these processes in the multilepton final state is challenging due to undetected neutrinos and the combinatorial ambiguity arising from multiple decay products. For the reconstruction and classification, we plan to exploit both permutation symmetry among final-state objects and Lorentz symmetry of their kinematics. We will utilize Lorentz-Equivariant Geometric Algebra transformers (LGATr)[1], whose layers are Lorentz-equivariant by construction and operate on multivectors of spacetime algebra, with controlled symmetry breaking introduced to match LHC physics. We will also test Lorentz Local Canonicalization (LLoCa)[2], which achieves equivariance through per-particle local frames and can be applied to arbitrary architectures. These approaches will be compared with standard transformer architectures. As Lorentz-equivariant networks benefit from input features that respect the underlying Lorentz symmetry[3], a key challenge is determining how best to represent the available kinematic information. The talk will present our approach and discuss the challenges and open questions in constructing meaningful representations for Lorentz-equivariant transformers

References:
1.arXiv:2405.14806
2.arXiv:2505.20280
3.arXiv:2608.02735

Orateur: JINU RAJ

18:30 → 22:30 Social Diner

mer. 18 novembre

09:00 → 10:45 Event/Object Reconstruction and Identification

Machine learning methods for reconstructing particles, events, and physical observables from detector data.

Examples include:
- Particle and event reconstruction
- Tracking and clustering
- Jet tagging and particle identification
- Multi-detector information fusion
- Reconstruction in high-granularity detectors

09:00 GammaLearn: deep-learning event reconstruction for the CTAO - Results and Lessons learned between 2017 and 2026

Since 2017, GammaLearn has developed deep-learning event reconstruction for Imaging Atmospheric Cherenkov Telescopes. Our main target is LST-1, the first Large-Sized Telescope of CTAO. One multi-task convolutional network, γ-PhysNet, estimates the particle type, energy and arrival direction of each event from the calibrated camera image and pixel timing.

From indexed convolution on hexagonal pixels to unsupervised domain adaptation, from Monte-Carlo to real data application, and from mono-telescope to stereoscopic reconstruction, this talk covers what we tried over eight years, what held up on real observations, and what is still open.

Orateur: Dr Thomas Vuillaume (LAPP, Univ. Savoie Mont-Blanc, CNRS)

09:20 Machine Learning for event reconstruction with uncertainty quantification in the ClearMind ultra-fast detector

The ClearMind collaboration [1] aims to develop a gamma detector with a coincidence resolving time below 100 ps. The detector combines a large monolithic crystal with a photocathode deposited directly on its surface and coupled to a microchannel plate photomultiplier tube (MCP-PMT). The opposite face is instrumented with silicon photomultipliers (SiPMs) to improve depth of interaction (DOI) and timing reconstruction. We develop a deep learning approach to reconstruct the 3D interaction position, deposited energy and interaction time from the detector signals, using events simulated with Geant4.
Building on previous work on density neural networks [2], we use the full waveforms to predict a probability density for each reconstructed parameter, providing both an estimate and its associated uncertainty. A two-dimensional convolutional neural network predicts the parameters of truncated Gaussian distributions for X, Y and DOI (truncated by the crystal dimensions), and an exponentially modified Gaussian distribution for interaction time. The network was trained on 700,000 simulated events depositing 511 keV in a 60 × 60 × 5 mm³ crystal.
The DNN reconstruction method achieves FWHM resolutions of 2.1 mm, 1.3 mm and 1.7 mm for X, Y and DOI, respectively. Predicted XY position uncertainties increase within 3 mm of the lateral borders, while DOI uncertainty decreases near the MCP face and, to a lesser extent, near the SiPM face. The obtained timing resolution is 131 ps FWHM, to be compared to 222 ps and 202 ps using constant fraction discrimination (CFD) on the first detected MCP and SiPM signals, respectively, and 160 ps when combining both. Selecting, for each parameter, the 20% of events with the lowest predicted uncertainty improves the respective X, Y, DOI and timing resolutions to 1.7 mm, 1.0 mm, 0.8 mm and 100 ps. We observe that the predicted timing uncertainty is correlated to the detected charge on both detector faces, proximity to the border, and number of Cherenkovs detected.
[1] Yvon, D., V. Sharyy, M. Follin, et al. « Design study of a “scintronic” crystal targeting tens of picoseconds time resolution for gamma ray imaging: the ClearMind detector ». Journal of Instrumentation 15, nᵒ 07 (2020): P07029‑P07029. https://doi.org/10.1088/1748-0221/15/07/P07029.
[2] Daniel, G., M. B. Yahiaoui, C. Comtat, et al. « Deep learning reconstruction with uncertainty estimation for γ photon interaction in fast scintillator detectors ». Engineering Applications of Artificial Intelligence 131 (mai 2024): 107876. https://doi.org/10.1016/j.engappai.2024.107876.

Orateur: Dr Théo Kaprélian (Université Paris-Saclay, CEA, Service de Génie Logiciel pour la Simulation, Gif-surYvette, 91191, Franc)

09:40 Transformers for reconstruction of electromagnetic showers in calorimeters

The reconstruction of electrons and photons in the CMS Electromagnetic Calorimeter (ECAL) currently relies on a geometrical clustering algorithm called PFClustering. While it is efficient for isolated particles, it has a limited ability to resolve close-by showers and mitigate detector noise, which reduces the sensitivity of physics analyses and will worsen with detector ageing. We present ClusTEX [1], a single-step graph-attention-based transformer that reconstructs photon energy and position directly from calorimeter readout. Through a novel positional encoding scheme, ClusTEX remains sensitive to the relative graph structure of the events while simultaneously learning detector-position dependencies. We demonstrate that this approach eliminates the need for a dedicated classification step and remains robust to complex event topologies and non-responsive detector regions. Trained and evaluated on a toy ECAL simulation, ClusTEX improves upon both PFClustering and GNN-based approaches. Particularly for events with overlapping showers, ClusTEX significantly outperforms PFClustering in signal efficiency and per-photon energy resolution. The impact of this improvement is reflected in the reconstruction of boosted $\pi^0\rightarrow\gamma\gamma$ decays, where ClusTEX successfully resolves pions above 30 GeV that are inaccessible to PFClustering. Beyond boosted $\pi^0$ reconstruction, this approach provides a promising foundation for jet reconstruction and improved prompt-photon identification in $\gamma+\text{jet}$ analyses. We report on the ongoing integration of ClusTEX into the CMS Software (CMSSW) and present first results using the full detector simulation for the LHC Run 3.

[1] Maidannyk, Y., Couderc, F., Malclès, J. and Sahin, M. Ö. Reconstruction of overlapping electromagnetic showers in calorimeters using Transformers. Eur. Phys. J. C 86, 873 (2026). https://doi.org/10.1140/epjc/s10052-026-16097-x

Orateur: Yuliia Maidannyk (CEA-IRFU)

10:00 Upgrade II LHCb calorimeter real-time reconstruction with GNN

Upgrade II constitutes a major leap forward at LHCb towards more and better data, driven by new hardware and software that will enable an ambitious increase of the current instant luminosity, thus, the experiment's physics reach. In particular, the calorimeter will be replaced by the new branded PicoCal, featuring better granularity and timing capabilities.
A significant challenge will be to keep or even improve accuracy with a much busier and granular detector, facing a huge increase of the data volumes to process real-time (30 MHz of inelastic collisions rate at LHCb). Studies by various groups among different institutes and experiments have demonstrated that GNNs can reach much better resolution than traditional clustering methods [1-4]. An important effort has been placed in making these architectures light enough so that they can run within the computing budget of the extreme data flow environments of real-time data processing in HEP experiments.
This presentation will show the preliminary results within the ODISSEE project to implement PicoCal GNN reconstruction in the Allen, LHCb's GPU High-Level-Trigger framework, and discuss the different GNN architectures that we are exploring to achieve the most performing results, looking forward to exchanging ideas with the comunity.

[1] Learning representations of irregular particle-detector geometry with distance-weighted graph networks
[2] Towards lightweight and distilled graph neural networks for efficient particle reconstruction in LHCb's next-generation calorimeter
[3] Graph Neural Network–based Energy Reconstruction for the LHCb Upgrade II Electromagnetic Calorimeter
4: Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics

Orateur: Aniol Lobo Salvia (CNRS - LAPP)

10:20 SPINE for DUNE

SPINE (Scalable Particle Imaging with Neural Embeddings) is a transformative, ML-native approach to reconstruction and is rapidly becoming a common reconstruction pipeline across LArTPC experiments. Building on its strong performance in ICARUS, SBND and ND-LAr, we are extending SPINE to the DUNE Far Detector and its two large-scale prototypes at CERN. This is not a direct port of existing configurations. The scale, readout geometries and operating conditions of these detectors call for substantial development at every level of the chain, from data preparation and training infrastructure to network design and the evaluation of reconstruction performance.

Using sparse CNNs and GNNs, SPINE performs end-to-end event reconstruction in one coherent framework, combining semantic segmentation, vertexing, particle clustering and kinematic reconstruction. We are adapting and improving each of these stages for the ProtoDUNE detectors and the Far Detector. We evaluate their performance with calibration-level benchmarks such as Michel electron tagging, electromagnetic shower reconstruction and dE/dx resolution across detector configurations. We are also developing a more efficient staged training workflow.

We also study cross-detector generalization and domain adaptation, which are essential for deploying SPINE at scale throughout the DUNE program.

Together, these developments move SPINE beyond its original deployments toward a scalable, experiment-agnostic reconstruction framework that can meet the demands of next-generation neutrino physics.

Orateur: Nived PUTHUMANA MELEPPATTU (IJCLab)

10:40 A machine learning approach to stellar stream detection in photometric surveys

Stellar streams are faint remnants of disrupted dwarf galaxies or globular clusters, orbiting within our Galaxy's halo. These low-surface-brightness structures can be detected in photometric surveys such as DES and LSST, but detection has so far relied on visual inspection. We aim to automate this process using a machine learning approach, trained on simulated streams injected with the StreamObs package.

Orateur: Matthieu Pelissier (LPSC/UGA)

10:45 → 11:15 ☕

11:15 → 12:30 AI at IN2P3/CNRS and at CEA/Irfu

11:15 Stratégie IA à l'IN2P3 - vers un GDR

Orateur: Julien DONINI (UBP/LPC/IN2P3)

12:30 → 14:00 Lunch

14:00 → 15:25 Foundation Models, Representation Learning and Generative AI

General-purpose learning paradigms for scientific data and physics applications.

Examples include:
- Self-supervised learning
- Foundation models
- Representation learning
- Transformers
- Graph representation learning
- Generative AI and diffusion models
- Physics-informed learning

14:00 Performance of Tabular Foundation Models for analysis of 4-vectors LHC final states with the FAIR Universe HiggsML Dataset

In collider-based particle physics experiments, proton collision events are commonly represented as tabular datasets for specific final states, features being four-vectors of final state particles or higher-level variables (invariant masses).
Inspired by the success of foundation models in language and vision, recent developments have introduced tabular foundation models such as TabPFN and TabICL.
We explore the use of tabular foundation models in collider physics analyses, and we show that they clearly outperform baselines like XGboost in data-limited regimes, for 10K training events or less. PFNs are benchmarked against boosted decision trees and neural networks across classification, density ratio evaluation (including calibration), and using the FAIR Universe dataset. We discuss performance, data efficiency, and computational trade-offs, and assess the potential role of pretrained tabular models in HEP analysis workflows.

Orateur: David Rousseau (IJCLab, Université Paris-Saclay)

14:20 CaloFound: towards a foundation model for calorimetry

CaloFound uses conditional shower generation to pretrain a reusable transformer representation in an unsupervised manner for calorimeter representation and reconstruction. Using GEANT4 simulations of a CMS-like electromagnetic calorimeter, it masks a ($9\times9$) crystal window around a photon impact, randomly placing its impact crystal within the central ($3\times3$) region. The encoder processes the visible detector, while a stochastic decoder generates the missing deposits conditioned on the incident energies and positions of all photons entering the window. Particle conditions enter only the decoder, allowing encoder reuse without generator-level information.
Distributional objectives constrain crystal energies, spatial energy profiles and hit occupancy across isolated, overlapping and spatially separated showers. The generated shower profiles agree with the reference simulation to within 2%.
We will present generation performance and transfer of the frozen and fine-tuned encoder to photon energy, position and seed-probability prediction. Comparisons with the state-of-the-art will assess energy and position resolution and the reconstruction of overlapping photons.

Orateur: Mehmet Ozgur Sahin (CEA-Saclay)

14:40 The Limits of Foundation Models for Jet Physics: Tagging, Generation, Scaling and Anomalies

Jets are ubiquitous but highly complex objects at the LHC, and learning their substructure patterns provides information about the particles that produced them and, ultimately, the underlying physics. They are therefore paramount to the LHC physics program. Foundation models have emerged as a powerful new ML paradigm for modelling jets and harnessing the information encoded in their constituents. In this talk, we will investigate how far foundation models can go toward learning and exploiting the complete information content of jets. We will approach this question from the perspective of the fundamental limits of jet tagging and generation, and their prospective scaling laws. Furthermore, we will explore the applicability of these representations to fully agnostic anomaly detection.

Orateur: Humberto Reyes-Gonzalez (RWTH Aachen)

15:00 Improving neural generative models with density estimation and Monte-Carlo resampling

Detector simulation is among the most resource-intensive components of modern collider experiments, currently consuming roughly half of the LHC computing budget and more still in the High-Luminosity phase. Normalizing flows are attractive surrogate models for fast simulation: they sit on the Pareto frontier between inference speed and accuracy compared with competing generators such as VAEs and GANs, and support both generation and tractable density estimation. Nevertheless, closing the measurable fidelity gap—especially at high detector granularity—is necessary for flows to be truly competitive. In this study, we improve the generation quality of normalizing flows without modifying their architecture.

First, we show that a classifier trained to distinguish reference from synthetic showers provides an effective estimator of the density ratio between the two distributions. Using this ratio with acceptance–rejection sampling, we achieve better agreement with the reference distribution across a range of physically motivated high-level observables. This generic technique can be used in principle to improve any generative model.

Second, exploiting the bijective nature of flows, we perform the same resampling directly in latent space, where it is much cheaper than in data space. A classifier-based density-ratio estimate between the latent and base distributions is used to resample the base distribution before the flow's backward transformation, achieving comparable fidelity gains with minimal additional time and compute.

Orateur: Tuan Pham (IJCLab - Paris Saclay)

15:20 FoSFor : Multiscale Graph Representations from Forest-Based Coarse-Graining

Representing large and complex scientific data as graphs requires methods that capture their structure across different scales while remaining computationally tractable. In this context, the FoSFor (Forest Screening for Multiscales Network Processing) toolbox, a CNRS Innovation Open project, provides a forest-based coarse-graining framework for constructing compact and multiscale graph representations.

The coarse-graining is driven by the stochastic dynamics associated with the graph rather than by purely geometric or topological criteria. At each step, a subset of representative states is selected through a random forest construction, and the dynamics is projected onto this reduced set while retaining a link with the original state space through an intertwining kernel.

Repeating this procedure produces a hierarchy of graphs and associated transition kernels across scales. The resulting representation is therefore not only a compact encoding of the graph, but also a dynamics-aware description in which coarse states retain information about the "microscopic" configurations they represent.
This provides a natural framework for studying graph signals at multiple scales and for constructing bases adapted to the underlying graph dynamics.

At each scale, the intertwining operator links the coarse representation to the original graph.
This construction naturally leads to a multiscale wavelet basis for graph signals, with approximation and detail functions associated with different resolutions. The resulting basis enables the analysis and reconstruction of signals at successive resolutions while retaining the link with the underlying graph dynamics.

This compact and structured representation can provide an efficient input for Graph ML and representation learning, while its multiscale structure supports graph signal processing and reconstruction tasks, including signal inpainting. For example, this approach achieves a recovery score of 0.8 when reconstructing a signal from 50% of the available signal information.

These methods are made available through the FoSFor toolbox, providing open-source implementations, pre-built binaries, and a user-friendly API to facilitate their adoption and integration into scientific workflows.

Orateurs: Dominique BENIELLI (Aix Marseille Université), Dr Sacha Duverger (Aix Marseille Univ, CNRS, I2M, Marseille, France)

15:25 → 15:35 Closing

15:25 Closing