Reconstruction of overlapping electromagnetic showers in calorimeters using Transformers

Yuliia Maidannyk, Fabrice Couderc, Julie Malclès, Mehmet Özgür Sahin · 2026

Read Reconstruction of overlapping electromagnetic showers in calorimeters using Transformers by Yuliia Maidannyk, Fabrice Couderc, Julie Malclès, Mehmet Özgür Sahin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-18172.txt.

Introduction

Maidannyk and co-authors use transformers to sort out showers that overlap in calorimeters. Pileup and boosted decays blur the energy blobs, so photon and electron calls get hard. Their ClusTEX graph transformer picks seeds and rebuilds energy and place in one step. A new code for place splits local graph terms from global detector terms. On GEANT4 tests, it sharpens energy and place calls, cuts fake splits, and holds up when some channels die.

Catalog record

CREATORYuliia Maidannyk, Fabrice Couderc, Julie Malclès, Mehmet Özgür Sahin
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69858
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.18172

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Reconstruction of overlapping electromagnetic showers in calorimeters using Transformers Yuliia Maidannyk, Fabrice Couderc, Julie Malclès, Mehmet Özgür Sahin Abstract Accurate clustering of electromagnetic energy deposits is essential for reconstructing photons and electrons in modern hadron collider experiments, where boosted topologies and pileup often cause overlapping showers and ambiguous energy assignment. We present deep learning-based clustering approaches that reconstruct particle energy and impact position directly from calorimeter readout. The study includes a two-step strategy in which candidate seed windows are identified and then jointly processed via distance-weighted message passing or attention mechanism, and a single-step graph transformer, ClusTEX , which performs candidate selection and reconstruction in one inference stage. ClusTEX uses a novel positional encoding scheme that separates local coordinates within the graph from global detector coordinates, enabling efficient, geometry-aware inference. Models are trained on Geant4 simulations of a simplified toy calorimeter and an ECAL-inspired topology with an explicit ( η , ϕ ) (\eta,\phi) dependence. Performance is evaluated using efficiency, energy and position resolutions, background rejection, and “splitting rate” — the probability to reconstruct two objects for a single-photon shower. In the toy calorimeter, attention-based interactions improve the reconstruction of overlapping showers relative to both the standard algorithm and distance-driven message passing, while maintaining performance on isolated photons and reducing splitting without multi-pass inference. In boosted π 0 → γ ​ γ \pi^{0}\to\gamma\gamma events, the attention-based model retains di-photon mass reconstruction capability, where the standard algorithm becomes inefficient. In the ECAL-inspired topology, ClusTEX provides the best overall performance, yielding improved energy resolution and reduced splitting compared to two-step approaches and the standard algorithm. ClusTEX also remains robust under localized detector failures, showing improved stability and partial recovery of energy in non-responsive channels. keywords ECAL, transformers, EM object reconstruction, photon, pion 1 Introduction In modern hadron collider experiments with a large number of simultaneous collisions and highly granular detectors, reconstructing the properties of individual particles from detector information is a major computational and algorithmic challenge. Electromagnetic calorimeters record the energy deposits left by particles and play a central role in the measurement of electron and photon four-momenta. To reconstruct the energy and position of the initiating particle, energy deposits in many detector cells must be clustered together, often in the presence of overlapping showers, detector nonuniformities, and experimental noise. These aspects can be illustrated using the Electromagnetic CALorimeter (ECAL) of the Compact Muon Solenoid (CMS) detector as an example. The CMS experiment is a general purpose detector located at the Large Hadron Collider (LHC) CMS Collaboration 2008 at CERN. Its physics programme includes precision studies of the Standard Model and searches for physics beyond it. These goals require the reliable identification and reconstruction of particles produced in proton-proton collisions at center of mass energies up to 13.6 TeV. In the High Luminosity LHC (HL-LHC) era, the delivered dataset and the number of additional simultaneous collisions (pileup) per bunch crossing will increase significantly, motivating the Phase-2 upgrade of the CMS detector and its reconstruction algorithms to sustain performance under these conditions. A large number of physics analyses performed with CMS data, including studies of Higgs boson properties CMS Collaboration 2023 ; CMS Collaboration 2020 , rely on a precise reconstruction of photons and electrons. In CMS, this reconstruction is carried out through a multi-step workflow, starting from localized energy deposits in the calorimeters and culminating in particle identification and origin assignment within the particle flow framework CMS Collaboration 2021 .The reconstruction of the kinematic properties of photons is implemented in the Particle Flow clustering algorithm, PFClustering CMS Collaboration 2017 . While this method is robust and efficient, it has limitations that become increasingly relevant in dense environments, at low energy, or for light boosted objects decays. In particular: 1. The increase in pileup expected at the HL-LHC will lead to a higher occupancy in the calorimeters meaning clustering ambiguities will become more frequent and the probability of incorrect associations between deposits and initiating particles will increase. 2. Its ability to separate nearby electromagnetic showers is limited. This affects the reconstruction of close-by photons, such as those produced in neutral pion ( \pi_{0} ) decays, which can mimic the signature of an isolated photon. Similar topologies also appear in searches for new phenomena, for instance the exotic decay H\rightarrow AA\rightarrow 4\gamma CMS Collaboration 2023 , where A denotes a light scalar or pseudoscalar particle. In order to cope with such signatures, some searches have employed end-to-end regression networks that directly target the final state observable (for example the reconstructed mass). While effective for a specific topology, such approaches are intrinsically specialized and cannot be straightforwardly reused across other physics analyses and search channels. 3. The algorithm exhibits a high background rate at very low energies (below approximately 1 GeV) related to detector noise. Such backgrounds are expected to increase with detector aging and can further degrade reconstruction performance. In addition to these aspects, the reconstruction of converted photons constitutes a particularly challenging regime. Photon conversions modify the observed topology through additional physics objects and the associated broadening of the spatial distribution of energy deposits, increasing the probability of overlaps and making the association of deposits to the correct initiating particle more difficult and ambiguous due to similar topologies associated to decays of other particles, and/or pileup. Improvements in the clustering stage can therefore directly benefit the reconstruction of converted photons, in addition to the gains obtained for unconverted photons. To address these limitations and improve the reconstruction of particle properties, machine learning (ML) methods can be considered. ML techniques are already widely used in analyses involving photons and electrons (for example Belayneh, D. et al. 2020 ; Akchurin, N. ; Cowden, C. ; Damgov, J. ; Hussain, A. ; Kunori, S. 2021 ) and even in the future detector design optimizations (see for example Schmidt et al. 2025 ; Nguyen et al. 2026 ). However, ML has not been adopted widely for energy clustering in electromagnetic calorimeters in a way that is optimized for multiple overlapping deposits originating from several particles, where new challenges emerge from high occupancy and detector-specific effects. Earlier work introduced the DeepCluster concept as a graph neural network (GNN) based approach to calorimeter clustering Simkina, P. and Couderc, F. and Malclès, J. and Sahin, M. Ö. 2024 . One shortcoming of the DeepCluster approach was the tendency to reconstruct two clusters for a single photon energy deposit due to enriched training on samples with two close-by photons. This issue was classified as ‘splitting’ and required multi-pass inference to re-merge these clusters into a single cluster corresponding to a single photon energy deposit. In this paper, we make a substantial step beyond the earlier work along two complementary venues. First, we present an updated DeepCluster style approach in which the conventional graph message passing is replaced by an attention mechanism. This modification retains the graph representation but enables a more adaptive exchange of information between detector cells, which improves the performance of the GNN-based approach and removes the necessity for an ad hoc multi-pass inference. Second, motivated by the practical difficulty of deploying two-step clustering networks in real event conditions and in full reconstruction chains, we introduce a new single-step graph-based transformer architecture. This single-step formulation reduces the reconstruction to

Read the complete text