Point-cloud generative models for fast calorimeter simulation across particles and geometries

arXiv:2609.30403 · physics.ins-det, hep-ex, hep-ph · Submitted 2026-09-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Point-cloud generative models for fast calorimeter simulation across particles and geometries".

Title and authors: Kai: The authors are Thorsten Buss, Henry Day-Hall, Frank Gaede, Gregor Kasieczka, Katja Krüger, Anatolii Korol, Thomas Madlener, Peter McKeown and Martina Mozzanica and Lorenzo Valente >

Mira: It's interesting to see this collaboration between different groups focusing on the simulation side >

Lev: They’re trying to make these models work for real hardware by using a point cloud representation which is more flexible than fixed grids when dealing with complex shapes >

Kai: So, what they summarize in this Point-cloud generative models for fast calorimeter simulation across particles and geometries paper is that deep generative models can reproduce the energy deposits orders of magnitude faster than Geant4 >

Mira: It’s about moving beyond the traditional simulation approach to something much quicker for generating those complex shower patterns >

The paper's summary: Kai: So, what this paper actually summarizes is that they developed several specific models to tackle different simulation needs in this area >

Mira: They detail CaloClouds3 which generates photon electromagnetic showers and runs about one hundred twenty times faster than Geant4 on a single CPU > <ref:2609.30403#pg1,faster than Geant4 on a single CPU>

Lev: That’s a concrete speed-up, but what I wonder is how that fidelity compares to the traditional method when you look at the energy deposits themselves >

Kai: Well, they claim it maintains fidelity to Geant4 results for photon showers while delivering this significant speed increase >

Mira: Then there's CaloHadronic, which uses transformer attention to extend the point cloud diffusion approach for pion hadronic showers that span both the electromagnetic and hadronic calorimeters >

Lev: That sounds computationally intensive, but using transformer self-attention to capture structure between points suggests they are modeling those spatial relationships more effectively than just a simple sum of points >

The paper's improvements: Kai: The paper also highlights transfer learning as a way to adapt these models to new detectors, which is super useful for real experiments >

Mira: They suggest that cross-geometry transfer learning only needs two to three orders of magnitude fewer training showers while keeping the generative performance good >

Lev: That reduction in required training data is important because it means we don't need to generate as much synthetic data beforehand, which cuts down on the upfront cost of generating that training set >

Kai: They show that pre-training on synthetic geometries can cut the simulated data needed for the network by two to three orders of magnitude >

Mira: That sounds like a really practical improvement because it addresses one of the biggest hurdles in deploying these models to real, unseen detectors >

Conclusion: Kai: So, to wrap up, this paper on Point-cloud generative models for fast calorimeter simulation across particles and geometries shows that we can achieve significant speed-ups for both photon showers with CaloClouds3 and hadronic showers with CaloHadronic >

Mira: And they’ve also shown that transfer learning helps us adapt these models to new detector shapes using much less training data, which is a big step toward making this technology practical across different experimental setups >

Lev: For real hardware, the challenge remains fitting the overhead of these diffusion models onto existing infrastructure while maintaining that claimed speed-up and fidelity >

Kai: We've seen how they use transformer attention for hadronic showers and how they reduce training data needs through transfer learning in this work on Point-cloud generative models for fast calorimeter simulation across particles and geometries >

Mira: It really shows the potential of these deep generative models to speed up the energy deposit calculation significantly compared to Geant4 >

Lev: We should keep watching how they improve hadronic fidelity, because that’s where the most complex physics is happening in those showers >

Thorsten Buss, Henry Day-Hall, Frank Gaede, Gregor Kasieczka, Katja Krücer, Anatolii Korol, Thomas Madlener, Öter Peter McKeown

University of Hamburg Institute for Experimental Physics Department of Deutsches Elektronen-Synchrotron DESY CERN

physics.ins-det, hep-ex, hep-ph

Submitted: 2026-09-24

Updated: 2026-09-24

Comments: Proceedings of the 14th Large Hadron Collider Physics Conference (LHCP 2026), 18-22 May 2026, Paris, France. 6 pages, 4 figures

Journal ref: PoS LHCP2026 (2026) 064

DOI: 10.22323/1.541.0064

License: http://creativecommons.org/licenses/by/4.0/

The gist: Point-cloud generative models for fast calorimeter simulation across particles and geometries summarize recent progress in generative point cloud fast simulation for highly granular calorimeters.

Key concepts

Point Cloud Models
These models represent particle showers as a set of energy deposits in 3D space, similar to a collection of points. This representation is chosen because it does not require pre-defined readout segmentation, making it suitable for calorimeters with complex or unknown geometries.
Fixed-Grid Models
These models tie the simulation directly to the segmentation of a specific detector. They are used when the detector's structure is known and segmented, but they are less flexible than point cloud models when dealing with complex or novel detector shapes.
CaloClouds3
This specific generative model is designed to simulate photon (electromagnetic) showers. It achieves a significant speed-up, running approximately 120 times faster than Geant4 simulations on a single CPU while maintaining accuracy for energy deposit calculations.
Transfer Learning
This technique allows models trained on one set of geometries or particle types to be adapted to new detectors. The paper shows that pre-training on synthetic data can reduce the required training showers by two to three orders of magnitude when adapting the model to a new geometry.

Terminology

Summary

Point-cloud generative models for fast calorimeter simulation across particles and geometries summarize recent progress in generative point cloud fast simulation for highly granular calorimeters. p. 0

Where the paper says it

de Detailed Geant4 simulation of calorimeter showers is the largest single computing cost of collider experiments, and the High-Luminosity LHC will need about ten times more simulated events than are currently produced. We summarise recent progress in generative point cloud fast simulation for highly granular calorimeters.

Deep generative models could reproduce the energy deposits orders of magnitude faster than Geant4. p. 1
Where the paper says it

The High-Luminosity LHC will need about ten times more simulated events. Deep generative models could reproduce the energy deposits orders of magnitude faster than Geant4.

Model Architectures Reviewed

The paper reviews four lines of work in point cloud generative fast simulation, ranging from a single particle type in one detector to twelve particle types in one compact model. The models fall into two broad families based on their native shower representation: fixed-grid models tied to the segmentation of one particular detector, and point cloud models which describe a shower as a set of energy deposits whose number varies from shower to shower. p. 1

Where the paper says it

Deep generative models could reproduce the energy deposits orders of magnitude faster than Geant4. They fall into two broad families depending on their native shower representation: fixedgrid models, tied to the segmentation of one particular detector, and point cloud models, which describe a shower as a set of energy deposits whose number varies from shower to shower [2, 3].

The authors adopt a point cloud representation because it presupposes no readout segmentation, allowing the description of detectors with complex geometries. p. 1
Where the paper says it

A point cloud presupposes no readout segmentation, so it can describe detectors with complex geometries, although the deposits themselves are still shaped by the detector that produces the shower. We therefore adopt a point cloud representation.

Specific Generative Models

Several models are detailed in the contribution, each addressing different simulation needs.

  1. CaloClouds3 generates photon (electromagnetic) showers and runs on average about 120× faster than Geant4 on a single CPU. p. 0

    Where the paper says it

    We summarise recent progress in generative point cloud fast simulation for highly granular calorimeters. CaloClouds3 generates photon (electromagnetic) showers, is geometry-independent, and runs on average about 120× faster than Geant4 on a single CPU.

  2. CaloHadronic uses transformer attention to extend the point cloud diffusion approach to pion (hadronic) showers spanning the electromagnetic and hadronic calorimeters. p. 0

    Where the paper says it

    CaloClouds3 generates photon (electromagnetic) showers, is geometry-independent, and runs on average about 120× faster than Geant4 on a single CPU. CaloHadronic uses transformer attention to extend the point cloud diffusion approach to pion (hadronic) showers spanning the electromagnetic and hadronic calorimeters.

  3. AllShowers uses a single model covering twelve particle types across both the electromagnetic and hadronic calorimeters. p. 2

    Where the paper says it

    CaloHadronic, like CaloClouds3, is specialised to one particle type, so covering all species this way would need one model per particle, which motivates a single model for all of them. AllShowers: one model for all particles AllShowers [6] is a single generative model covering twelve particle types across both the electromagnetic and hadronic calorimeters, namely 𝑒 ± , 𝛾, 𝜋 ± , 𝐾 ± , 𝐾 0 𝐿 , 𝑝, 𝑝¯, 𝑛 and 𝑛¯, where earlier models each covered only photons, electrons or 𝜋 + .

Performance Metrics and Speed-up

The performance gains of these models are quantified by comparing simulation time per event against Geant4.

CaloClouds3 is on average about 120× faster than Geant4 (Fig. 1)

Transfer Learning and Geometry Adaptation

Cross-geometry transfer learning is presented as a method to adapt models to new detectors.

Transfer learning needs two to three orders of magnitude fewer training showers while preserving the generative performance.

The authors demonstrate that pre-training on synthetic geometries can cut the simulated data needed to train the network by two to three orders of magnitude.

Future Directions

Open directions for future work include improving hadronic fidelity and extending the synthetic geometry prior to all particle species and towards zero-shot simulation of unseen detectors. The goal is to achieve zero-shot simulation of unseen detectors. p. 4

Where the paper says it

Transfer learning [7, 8] reaches a new geometry from very few target showers and cuts the simulated data needed to train the network by two to three orders of magnitude. Open directions are to improve the hadronic fidelity and to extend the synthetic geometry prior to all particle species and to the full calorimeter description, towards zero-shot simulation of unseen detectors.

Acknowledgements

The work utilized computational resources at DESY and received funding from the European Union’s Horizon 2020 Research and Innovation programme under Grant Agreement No 101004761. The research was supported by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy, EXC 2121 Quantum Universe. p. 4

Where the paper says it

It has received funding from the European Union’s Horizon 2020 Research and Innovation programme under Grant Agreement No 101004761, from the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy, EXC 2121 Quantum Universe, 390833306, and via the KISS consortium (05D23GU4, 13D22CH5) funded by the German Federal Ministry of Research, Technology and Space (BMFTR) in the ErUM-Data action plan. is supported by the Helmholtz Initiative and Networking Fund’s initiative for refugees as a refugee of the war in Ukraine.

A.K. is supported by the Helmholtz Initiative and Networking Fund’s initiative for refugees as a refugee of the war in Ukraine. p. 4
Where the paper says it

It has received funding from the European Union’s Horizon 2020 Research and Innovation programme under Grant Agreement No 101004761, from the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy, EXC 2121 Quantum Universe, 390833306, and via the KISS consortium (05D23GU4, 13D22CH5) funded by the German Federal Ministry of Research, Technology and Space (BMFTR) in the ErUM-Data action plan. is supported by the Helmholtz Initiative and Networking Fund’s initiative for refugees as a refugee of the war in Ukraine.

P.M. is supported by the CERN Strategic R&D Programme on Technologies for Future Experiments. p. 4
Where the paper says it

is supported by the Helmholtz Initiative and Networking Fund’s initiative for refugees as a refugee of the war in Ukraine. is supported by the CERN Strategic R&D Programme on Technologies for Future Experiments [12].

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

  1. Fast simulation of photon showers can be achieved by using CaloClouds3, which is about 120x faster than Geant4 on a single CPU. This allows for rapid energy deposit calculation in simulations without sacrificing fidelity to Geant4 results.

  2. Generation of hadronic showers across both electromagnetic and hadronic calorimeters can be accomplished with CaloHadronic, which uses transformer self-attention to capture structure between points, enabling holistic pion shower generation.

  3. A single model can cover twelve particle types in one compact model called AllShowers, which replaces specialized baselines and matches or exceeds their fidelity while requiring only a small number of trainable parameters.

  4. Cross-geometry transfer learning enables the application of models to new detectors by needing two to three orders of magnitude fewer training showers, significantly reducing the data required for geometry adaptation.

Abstract

Detailed Geant4 simulation of calorimeter showers is the largest single computing cost of collider experiments, and the High-Luminosity LHC will need about ten times more simulated events than are currently produced. We summarise recent progress in generative point cloud fast simulation for highly granular calorimeters. CaloClouds3 generates photon (electromagnetic) showers, is geometry-independent, and runs on average about 120x faster than Geant4 on a single CPU. CaloHadronic uses transformer attention to extend the point cloud diffusion approach to pion (hadronic) showers spanning the electromagnetic and hadronic calorimeters. AllShowers unifies twelve particle types in a single model with far fewer parameters than the specialised baselines while matching or exceeding their fidelity. We close with cross-geometry transfer learning, which needs two to three orders of magnitude fewer training showers while preserving the generative performance.

Sources

Related papers