3D software compensation of hadronic showers in the CRILIN crystal calorimeter
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "3D software compensation of hadronic showers in the CRILIN crystal calorimeter".
Mira: Here's the gist of the paper. Future electron–positron colliders such as FCC-ee require excellent jet energy resolution to fully exploit their Higgs physics potential and to precisely measure Higgs boson couplings to quarks.
Kai: In this work, software compensation techniques are investigated for CRILIN, a crystal ECAL based on a semi-homogeneous concept, featuring longitudinally segmented Cherenkov PbF2 crystal matrices and silicon photomultiplier (SiPM) readout.
Mira: The ParticleNet Graph Neural Network exploiting the full three-dimensional shower topology achieves significantly improved performance with respect to simple energy sum reconstruction.
Kai: The reconstruction of hadronic showers is fundamentally complicated by fluctuations of the electromagnetic fraction fEM generated by neutral pion production.
Paper summary: Kai: To summarize this paper on "three dee software compensation of hadronic showers in the CRILIN crystal calorimeter", they're looking at how to boost jet energy resolution for future colliders <ref:2606.05111#pg1,3D software compensation of hadronic showers in the CRILIN crystal calorimeter>. The main idea is that because these crystal calorimeters have a very non-compensating response, you need better ways to reconstruct the hadron energy.
Mira: They are testing software compensation techniques on CRILIN, which uses longitudinally segmented Cherenkov PbF2 crystals with silicon photomultiplier readout. The thesis is that they can use shower shape observables—like transverse RMS and longitudinal center of gravity—to correct for the energy loss due to these fluctuations.
Lev: It claims that even without using dual-readout for an intermediate fraction estimation, software compensation in this Cherenkov-based calorimeter with three dimensional readout can reach the performance level of a dual-readout calorimeter when using Graph Neural Networks.
Kai: The paper specifically points out that crystal calorimeters are a tough case because their great electromagnetic performance comes at the cost of a highly non-compensating response, which could really hurt the combined ECAL plus HCAL resolution if you put an HCAL with good resolution downstream.
Mira: They suggest that modern calorimeters with fine longitudinal and transverse segmentation have detailed information about how the shower develops, and that this detailed information can be used to recover some of what's lost.
Lev: The method they use, the ParticleNet architecture for the GNN, takes those hits—position and energy—and it learns to map them to the true energy deposited inside CRILIN.
Conclusion: Kai: Looking at the work by Ciccarella, Gargiulo, Sarra, et al., about this three dee software compensation for CRILIN, the implication is that this type of detector isn't totally hopeless when it comes to hadronic energy reconstruction <ref:2606.05111#pg1>.
Mira: They show that even though crystal calorimeters are inherently non-compensating, using these software techniques can recover a lot of the information lost from shower fluctuations.
Lev: So what this means practically is that this technique could be a viable alternative to needing dual-readout calorimetry in ECALs for achieving the necessary resolution targets at future colliders.
Kai: They found that CRILIN contributes less than five percent to the total hadronic energy resolution for beam energies above thirty GeV, which is a solid number for this type of detector <ref:2606.05111#pg2>.
Mira: And they also say that for beam energies above forty GeV, the contribution of CRILIN to the combined calorimetric resolution is below approximately five percent.
Lev: It’s important to remember that the method they used relies on those detailed shower development features being available from their segmentation and readout system.
Kai: So, this paper shows that detailed information about how a shower develops can help correct for non-compensation in these crystal detectors by using advanced neural network techniques.
Laboratori Nazionali di Frascati dell’INFN Sezione di Frascati, Italy · INFN Sezione di Firenze, Florence, Italy · INFN Sezione di Roma1, Rome, Italy
hep-ex
Submitted: 2026-06-03
Updated: 2026-11-16
Comments: Version accepted by JINST
Journal ref: 2026 JINST 21 P11018
DOI: 10.1088/1748-0221/21/11/P11018
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 87/100
The gist: Future electron–positron colliders such as FCC-ee require excellent jet energy resolution to fully exploit their Higgs physics potential and to precisely measure Higgs boson couplings to quarks.
Key concepts
- Non-compensating Calorimeters
- These detectors respond differently to electromagnetic and hadronic energy deposits. This difference causes fluctuations in the electromagnetic fraction (fEM) from neutral pion production, leading to degraded energy resolution and a non-linear response in the measured energy.
- Transverse Shower RMS (RRMS)
- This is a measure of how spread out the shower is perpendicular to its axis. It quantifies the width of the shower by calculating the root mean square distance between each hit and the center-of-gravity, providing a simple observable for energy correction.
- ParticleNet Graph Neural Network (GNN)
- A sophisticated AI model used here that treats calorimeter hits as nodes in a graph. Each hit is represented by its 3D position and energy. The network learns complex relationships across the entire shower topology to reconstruct the true deposited energy, outperforming simpler methods.
- Stochastic Term (S·√E[GeV])
- This term represents the statistical uncertainty in the reconstructed energy measurement. The paper shows that by applying software compensation, this uncertainty is drastically reduced—for example, from 69.4% before correction to 27.6% after using the GNN.
Terminology
Summary
Future electron–positron colliders such as FCC-ee require excellent jet energy resolution to fully exploit their Higgs physics potential and to precisely measure Higgs boson couplings to quarks. p. 1 In this work, software compensation techniques are investigated for CRILIN, a crystal ECAL based on a semi-homogeneous concept, featuring longitudinally segmented Cherenkov PbF2 crystal matrices and silicon photomultiplier (SiPM) readout. p. 2 The ParticleNet Graph Neural Network exploiting the full three-dimensional shower topology achieves significantly improved performance with respect to simple energy sum reconstruction. p. 0
Motivation and Context
The reconstruction of hadronic showers is fundamentally complicated by fluctuations of the electromagnetic fraction fEM generated by neutral pion production. In non-compensating calorimeters, where the response to electromagnetic and hadronic energy deposits differs, these fluctuations lead to degraded energy resolution and non-linear response. p. 1 Crystal calorimeters present a particularly challenging case because their excellent electromagnetic performance comes at the price of a highly non-compensating response. p. 2 Modern calorimeters featuring fine longitudinal and transverse segmentation contain detailed information about shower development that can potentially be exploited to recover part of the lost information. p. 2
Simple Software Corrections
Two approaches for software compensation were investigated for CRILIN. The first approach involves simple one-dimensional corrections based on shower-shape observables, such as cluster RMS width and longitudinal cluster center-of-gravity. p. 2
** Transverse shower RMS is defined as RRMS = √Í i Ei ®r i − ®r cog2 Í i Ei. p. 5 **
The primary quantity used throughout this work to derive energy corrections is the ratio E reco/E true between the reconstructed energy in CRILIN, i. p. 5e., the simple sum of the hits, and the true energy deposited inside CRILIN. p. 5 This correction is applied as Ecorr = Ereco / (p0 + p1RRMS). p. 8
Advanced Neural Network Compensation
The second approach exploits the full three-dimensional shower topology using a Graph Neural Network, taking as inputs the calorimetric hits and yielding the corrected total reconstructed energy in CRILIN as an output. The ParticleNet architecture is based on a Graph Neural Network (GNN) developed for application in high-energy particle physics. p. 0
** The input graph is built from reconstructed calorimeter hits, where each hit is represented by four features, i. p. 2e., the three-dimensional position and the energy. p. 0 p. 0 p. 0 **
The network target is the true energy deposited inside CRILIN, E CRILIN true = Ebeam − EVD. p. 5
Performance and Results
Both simple one-dimensional corrections and the GNN approach show that the hadron energy resolutions of a combined ECAL+HCAL system significantly improves with respect to non-compensating energy reconstruction.
** The stochastic term (S·√E[GeV]) is reduced from (69. p. 7 p. 74 ± 2.8)% before the correction to (34. p. 64 ± 0.3)% after the correction with the GNN. p. 0 **
** The stochastic term (S·√E[GeV]) is reduced from (69.4 ± 2.8)% before the correction to (39. p. 63±0.3)% after the correction with the Z-based correction. p. 7 **
The GNN reconstruction yields a stochastic term of (27. p. 06 ± 0.8)% after the correction, which is better than that obtained for one-dimensional corrections. p. 2 p. 2
Conclusion on Contribution
The contribution of CRILIN to the combined hadronic energy resolution is below 5% for energies above 30 GeV. The results demonstrate that a Cherenkov crystal calorimeter, despite its intrinsically noncompensating response, can recover a large fraction of the information lost by hadronic shower fluctuations through software compensation techniques. p. 0 This suggests that the technique described could represent a viable alternative to dual-readout calorimetry in ECALs. p. 1 The contribution of CRILIN to the total calorimetric resolution is below approximately 5% for beam energies above 40 GeV. p. 3
A Studies on the Cherenkov response with a muon simulated sample
The ratio between the Cherenkov response, C, and the deposited energy, Edep, can then be written for muons as C mu + C delta / Edep = α ΔL / Edep + max0, βEdep − γ / Edep. This non-linear effect is observed for muons and is accounted for throughout the paper. p. 5
References
[1] W. Bartmann, J.-P. Burnet, C. Carli, A. Chance, P. Craievich, M. Giovannozzi, C. Grojean, J. Gutleber, K. Hanke, A Henriques, P Janot, C Lourenco et al., Future circular collider feasibility study report volume 1: Physics and experiments (2025). p. 0 doi:10.17181/CERN. p. 29DKX.TDH9.
[2] S Lai, J Utehs, A Wilhahn, O Bach et al., Dual readout with capillary tubes – status and prospects, EPJ Web of Conferences 320 (2025) 00054. p. 1 doi:10.1051/epjconf/202532000054.
[3] S Lee, M Livan, R Wigmans, Dual-readout calorimetry, Reviews of Modern Physics 90 (2) (Apr. p. 1 2018). doi:10.1103/revmodphys.90.025002.
[4] N Valle, Dual readout with capillary tubes – status and prospects, EPJ Web of Conferences 320 (2025) 00154. p. 1 doi:10.1051/epjconf/202532000054.
[5] N Akchurin, C Cowden, J Damgov, A Hussain et al., The (un)reasonable effectiveness of neural network in cherenkov calorimetry, Instruments 6 (4) (2022). p. 0 doi:10.3390/instruments6040043.
[6] S Ceravolo, F Colao, C Curatolo et al., Crilin: A crystal calorimeter with longitudinal information for a future muon collider, Journal of Instrumentation 17 (09) (2022) P09033. p. 1 doi:10.1088/1748-0221/17/09/P09033.
[7] C Cantone, A Cemmi, S Ceravolo et al., Crilin: a semi-homogeneous crystal calorimeter for the muon collider, in: Proceedings of 42nd International Conference on High Energy Physics — PoS(ICHEP2024), Vol. p. 2 p. 2 p. 2 p. 2 476, 2024, p. 1113. doi:10.22323/1.476.1113.
[8] R Gargiulo, C Cantone, A Cemmi et al. p. 0 p. 0, Crilin: a semi-homogeneous crystal calorimeter for the muon collider, in: Proceedings of 42nd International Conference on High Energy Physics — PoS(ICHEP2024), Vol. 476, 2024, p. 1113. doi:10.22323/1.476.1113.
[9] S Ceravolo, M Cavallina, V Ciccarella et al. p. 0 p. 0, Design and beam-test characterization of the crilin semi-homogeneous crystal calorimeter (2026). p. 2 p. 2 arXiv:2607. p. 0 p. 028477. URL https://arxiv. p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0org/abs/2607.28477.
[10] T Basaglia, Z W Bell, D D’Agostino et al., Geant4: a game changer in high energy physics and related applicative fields (2024). p. 2 p. 2 arXiv:2405. p. 0 p. 012159. URL https://arxiv.org/abs/2405.12159.
[11] A Pellecchia, et al., Design and optimization of a hadronic calorimeter based on micropattern gaseous detectors for a future experiment at the muon collider, Nuclear Instruments and Methods in Physics Research Section A (2024). p. 2 p. 2 arXiv:2407. p. 0 p. 010261. URL https://arxiv.org/abs/2407.10261.
[12] I Frank, I Tamm, Coherent Visible Radiation of Fast Electrons Passing Through Matter, Springer Berlin Heidelberg, Berlin, Heidelberg, 1991, pp. p. 4 29–35. doi:10.1007/978-3-642-74626-0 2.
[13] J E Gaiser, Charmonium spectroscopy from radiative decays of the j/ψ and ψ', Tech. rep., SLAC National Accelerator Laboratory (SLAC), Menlo Park, CA (United States) (08 1982). p. 2 doi:10.2172/1453988.
[14] I Sarra, Development and test of the crilin calorimeter for the muon collider: Status and perspectives, in: 2025 IEEE Nuclear Science Symposium (NSS), Medical Imaging Conference (MIC) and Room Temperature Semiconductor Detector Conference (RTSD), 2025, pp. 1–1. doi:10.1109/NSS/MIC/RTSD57106.2025.11287.
[15] D Garcia, L Herrmann, G Krzmanc et al., End-to-end event reconstruction for precision physics at future colliders (2026). p. 0 arXiv:2603. p. 004084. URL https://arxiv.org/abs/2603.04084.
[16] H Qu, L Gouskos, Jet tagging via particle clouds, Physical Review D 101 (5) (Mar. p. 1 2020). doi:10.1103/physrevd.101.056019.
[7] C Cantone, A Cemmi, S Ceravolo et al., Crilin: a semi-homogeneous crystal calorimeter for the muon collider, in: Proceedings of 42nd International Conference on High Energy Physics — PoS(ICHEP2024), Vol. 476, 2024, p. 1113. doi:10.22323/1.476.1113.
[8] R Gargiulo, C Cantone, A Cemmi et al., Crilin: a semi-homogeneous crystal calorimeter for the muon collider, in: Proceedings of 42nd International Conference on High Energy Physics — PoS(ICHEP2024), Vol. 476, 2024, p. 1113. doi:10.22323/1.476.1113.
[9] S Ceravolo, M Cavallina, V Ciccarella et al., Design and beam-test characterization of the crilin semi-homogeneous crystal calorimeter (2026). arXiv:2607.28477. URL https://arxiv.org/abs/2607.28477.
[10] T Basaglia, Z W Bell, D D’Agostino et al., Geant4: a game changer in high energy physics and related applicative fields (2024). arXiv:2405.12159. URL https://arxiv.org/abs/2405.12159.
[18] A Pellecchia, et al., Design and optimization of a hadronic calorimeter based on micropattern gaseous detectors for a future experiment at the muon collider, Nuclear Instruments and Methods in Physics Research Section A (2024). arXiv:2407.10261. URL https://arxiv.org/abs/2407.10261.
[A] I Sarra, Development and test of the crilin calorimeter for the muon collider: Status and perspectives, in: 2025 IEEE Nuclear Science Symposium (NSS), Medical Imaging Conference (MIC) and Room Temperature Semiconductor Detector Conference (RTSD), 2025, pp. 1–1. doi:10.1109/NSS/MIC/RTSD57106.2025.11287.
[B] Dual readout calorimetry [4]. p. 1
[C] Software compensation based on shower topology [2]. p. 1
[D] Transverse shower RMS. p. 0
[E] Longitudinal center-of-gravity. p. 0
[F] ParticleNet training and results. p. 1
[G] Input graph is built from reconstructed calorimeter hits, where each hit is represented by four features, i. p. 2e., the three-dimensional position and the energy.
[H] The network target is the true energy deposited inside CRILIN, E CRILIN true = Ebeam − EVD. p. 5
[I] The dependence of the result on the assumed HCAL resolution is also studied and found to be limited within the range considered. p. 0
[J] The noise term is (1. p. 302 ± 0.05) GeV, which is compatible with the one found using the other approaches. p. 2
[L] The contribution of CRILIN to the combined hadronic energy resolution is below 5% for energies above 30 GeV. p. 1
[M] To take this non-linear effect into account, throughout the paper, pions that do not start any hadronic shower in the CRILIN module (dubbed MIP events) are treated as muons, and the proper C/Edep function is employed to convert the Cherenkov response to an energy value. p. 5
[N] The noise term of approximately 1 GeV comes from the relatively large thresholds applied to single hits, combined with the fact that Cherenkov light readout collects only energy from relativistic particles. p. 7
[O] A global factor comprising both efficiencies is evaluated starting from the measured photoelectron (p. p. 4e.) light yield of the crystal–photosensor system. p. 4
[P] Poisson fluctuations corresponding to this light yield are applied on an event-by-event basis, and crystal hits below 40 MeV are discarded to emulate the hit-reconstruction threshold used in realistic detectors to suppress electronic noise. p. 4
[Q] This conversion factor has been cross-checked using a 5 GeV muon sample, for which the ratio between the number of Cherenkov photons and the deposited energy, Edep, is found not to be constant as a function of Edep. p. 5
[R] The dependence of the result on the assumed HCAL resolution is also studied and found to be limited within the range considered. p. 0
[S] Software compensation techniques for the event-by-event correction of reconstructed hadron energies in the longitudinally segmented Cherenkov crystal CRILIN calorimeter have been investigated. p. 0
[T] Simple corrections based on global shower observables such as the transverse shower RMS or the longitudinal shower center-of-gravity already provide substantial improvements with respect to the uncorrected response. p. 0
[U] A ParticleNet Graph Neural Network exploiting the full three-dimensional hit information provides a robust alternative, outperforming the one-dimensional corrections for beam energies. p. 9
How the training signal is built
The input graph is built from reconstructed calorimeter hits, where each hit is represented by four features, i.e., the three-dimensional position and the energy. The network architecture is based on ParticleNet, a Graph Neural Network (GNN) developed for application in high-energy particle physics. p. 0
** The network consists of three EdgeConv blocks with k = 50 nearest neighbours and feature dimensions of 32, 32 and 64 channels, respectively, followed by a global average pooling layer. p. 3 **
Training is performed using the Adam optimizer with learning rate equal to 10−4, batch size 64 and mean-squared-error loss. p. 2 The network is trained for 60 epochs on a Tesla T4 GPU, using 20% of the dataset for validation. p. 2
Where the gains come from
The GNN reconstruction yields a stochastic term (S·√E[GeV]) of (27.6 ± 0.8)% after the correction, which is better than that obtained for one-dimensional corrections. The larger constant term obtained with the GNN calls for further studies in which global observables such as the RMS are provided as additional inputs to the network after the ParticleNet block. p. 2 The contribution of CRILIN to the combined hadronic energy resolution is below 5% for beam energies above 40 GeV. p. 1
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Software compensation using shower observables can be used to correct reconstructed energy on an event-by-event basis, as shown by equations like Ecorr = Ereco p0 + p1RRMS. This allows for a substantial improvement in hadronic energy reconstruction compared to uncorrected responses.
-
A ParticleNet Graph Neural Network exploiting the full three-dimensional shower topology can achieve significantly improved performance with respect to simple energy sum reconstruction. This network takes reconstructed calorimeter hits as input and yields the corrected total reconstructed energy in CRILIN as an output.
-
The GNN-based reconstruction, when compared to one-dimensional corrections, shows a stochastic term (S · √E[GeV]) of 27.6 ± 0.8% after correction, which is significantly lower than the value obtained for one-dimensional corrections of (69.4 ± 2.8)%.
-
The system can be improved by implementing a combined resolution formula that incorporates CRILIN and HCAL contributions, showing that the contribution of CRILIN to the combined hadronic energy resolution is below 5% for beam energies above 40 GeV.
-
The AI system can estimate the true energy deposited inside CRILIN as ECRILIN true = Ebeam − EVD by using Geant4 simulations and knowledge of the beam and VD energy. This estimation is crucial for accurate software compensation techniques like those based on transverse shower RMS or longitudinal center-of-gravity.
Sources
- Software Compensation for Highly Granular Calorimeters using Machine Learning
- Design and beam-test characterization of the CRILIN semi-homogeneous crystal calorimeter
- Geant4: a Game Changer in High Energy Physics and Related Applicative Fields
- Design and optimization of a hadronic calorimeter based on micropattern gaseous detectors for a future experiment at the Muon Collider
- End-to-end event reconstruction for precision physics at future colliders
Related papers
- DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII
- CANTON- mu Proposal: A Next-Generation Muon g-2 Measurement at Sub-0.1 ppm Precision
- Search for dark photon dark matter near 1,meV using a superconducting mixer
- ANUBIS detector and its sensitivity to neutral long-lived particles
- How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
- Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques