Calorimeter showers are the most expensive part of detector simulation, and highly granular calorimeters, in which a single hadronic shower leaves thousands of energy deposits, make them more expensive still. We present CaloCHILL (Calorimeter Hits In Learned Latents), a generative model that does not produce these deposits one by one: instead it generates a compact summary of the shower in a learned latent space and expands it to full resolution only at the end. The complete shower, including the arrival time of every hit, is built in three stages: a flow-matching model first sets the number of hits and the earliest hit time in each calorimeter layer, a mean-flow model then generates the latent summary, and the decompressor of a variational autoencoder turns it into the full point cloud, giving every hit its position, energy and time. Sparse attention along the shower axis keeps this compression affordable for showers with thousands of hits. On positive-pion showers in the calorimeter of the International Large Detector, a proposed detector for the FCC-ee, CaloCHILL reproduces the main shower observables, including their time structure. This new model is 600 times faster than its predecessor, while improving on all accuracy metrics considered.
Figures & tables
Figure 1: Reduction of a Geant4 shower to a point cloud, illustrated for a single readout cell. Left: the energy deposits produced by Geant4 , a quasi-continuous set of steps of which only a few are drawn. Centre: the virtual grid of CaloHadronic [ 19 ] , about nine times finer than the physical readout; this is the representation used here. Right: the physical readout granularity, recovered at the analysis stage, where all deposits inside a cell are merged into a single hit.
Figure 2: Generation pipeline of CaloCHILL . Conditioned on the energy of the incident particle, a flow-matching model ( LayerFM ) first produces the per-layer observables: the number of points and the earliest hit time in each of the 78 layers. Together with the incident energy, these condition both stages that follow. Starting from noise, the mean-flow model generates a latent code, which the VAE decompressor expands into the full size point cloud carrying position, energy and time; the layer index of each point is attached last, completing the five features of every point.
Figure 3: Sparse attention masks: the VAE compressor compressing a cloud onto the latent (left), the latent self-attention (centre), and the VAE decompressor expanding it back (right). Rows are queries, columns keys and values, a filled entry is one that is evaluated, and the lighter regions are padding. The example is a toy one, 1515 points onto ds=512 , sized for legibility rather than matching a configuration used in production.
Figure 4: A π+ shower at E=50.2\text{,}\mathrm{GeV}$$ : Geant4 (left), its reconstruction by the VAE , encoded and decoded once (centre), and a shower generated by the large LMF at ds=256 (right) with the incident energy and the per-layer numbers of points and earliest times of the Geant4 shower as conditioning. The axes x and z are transverse and the depth y vertical; the size of a cell grows with the square root of its energy and its colour gives the time of its earliest hit, from the start of the shower (red-orange) to 7ns (dark indigo); later cells are grey. The grey plane is the boundary between the ECal and the HCal.
Figure 5: Shower observables at the readout granularity for Geant4 (grey), the five latent sizes (blue) and the large LMF at ds=128 , 256 and 512 (green), 100000 showers each between 10GeV and 90GeV . From left to right: longitudinal profile, radial profile, cell-energy spectrum and apparent-velocity profile (distance from the gun over arrival time). The dashed lines in the last panel are the largest apparent velocities ceff of an ECal and an HCal cell, the solid line the speed of light. The insets show the last ECal layers of the longitudinal profile and the two MIP peaks of the cell-energy spectrum on a linear scale. The bands are the statistical uncertainty, and the lower panels the ratio to Geant4 .
WD/σ\textscGeant4 [ ×10−2 ]
JSD [ ×10−4 ]
ds=64
3.98±0.18
12.77±0.32
ds=128
3.41±0.18
10.22±0.27
ds=256
3.65±0.11
10.66±0.59
ds=512
3.94±0.25
10.87±1.10
ds=1024
6.56±0.22
23.68±0.42
ds=128 , large LMF
3.88±0.51
10.65±0.73
Table 1: Geometric mean over the observables of the normalised Wasserstein distance (WD) and of the Jensen–Shannon divergence (JSD) between each latent size and Geant4 , on the full-range samples of 100000 showers, from table 7 . The best model is in bold; the error is the standard deviation of the mean over 1000 resamplings of every entry within its uncertainty, and the last row is the Geant4 floor.
Figure 6: Energy sum (left), mean energy per layer (centre) and mean energy per shower in bins of the distance from the shower axis (right) at 80GeV for Geant4 (grey), AllShowers with 32 and 4 network evaluations (NFE 32, NFE 4), CaloHadronic , ds=128 and the large LMF at ds=256 ; the insets show the last ECal layers (centre) and the first three radial bins (right) on a linear scale. The lower panels give the ratio to Geant4 . The same comparison at 20GeV and 50GeV is in figure 14 .
20 GeV
50 GeV
80 GeV
±
Mean
WD/σ\textscGeant4 [ ×10−2 ]
CaloHadronic
7.23
5.55
5.19
0.70
5.99 ± 0.34
AllShowers NFE 32
5.16
4.06
3.36
0.90
4.19 ± 0.40
AllShowers NFE 4
6.70
6.28
6.40
0.97
6.46 ± 0.43
ds=128
6.87
5.52
5.24
0.61
5.88 ± 0.27
ds=256 , large LMF
5.84
4.97
5.00
0.41
5.27 ± 0.21
Table 2: Geometric mean over the observables of the WD and of the JSD to Geant4 at 20 / 50 / 80GeV , and their mean over the three energies, from table 9 ; the best model is in bold, the column ± is the largest uncertainty of the three energies, and the last row is the Geant4 floor.
Figure 7: Generation time per shower against incident energy on an NVIDIA A100 ( 80GB , PCIe) GPU at batch size 1400, for the full chain at ds=128 , the large LMF at ds=256 on a VAE trained without the hit time (orange) and with it (green, nearly on top of the orange line), compared with AllShowers at 4, 8 and 32 network evaluations. The dotted line extends a straight-line fit between 10GeV and 90GeV to 120GeV , an upper estimate since the model is not trained above 90GeV .
Figure 8: Accuracy against generation time. Vertical axis: JSD to Geant4 of table 2 , the geometric mean over the observables averaged over 20, 50 and 80GeV , with its error; Geant4 against itself gives the statistical floor. Horizontal axis: time per shower on an NVIDIA A100 GPU from table 11 , at batch size 1 (filled markers) and at batch size 1400 (open markers; 256 for CaloHadronic , the fastest batch size measured for it); for Geant4 a cost-normalised CPU time is given (see appendix J ).
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Level
Stage
Parameters
Layer-level
LayerFM (single-step student)
754332
Point-level
VAE
1553496
LMF
1327967
Total
3635795
Appendix
Table 3: Number of trainable parameters of each stage, for the baseline ds=256 configuration. The layer-level stage is LayerFM ; the point-level stages are the VAE , VAE compressor and VAE decompressor together, and the LMF .
Parameter
Value
Preprocessing
points
dequantised, log(n+0.5) , std.
earliest hit time
quantile transform (1000)
incident energy
log , std.
layers
78
Model
Appendix
Table 4: Hyper-parameters of LayerFM . Each transformation acts on each of the 78 layers independently; std.: standardised.
Parameter
Value
Preprocessing
point features
x , z , energy, time
x , z
scaled to [−1,1]
energy, time
log , std.
Model
latent points ds
64, 128, 256, 512, 1024
Appendix
Table 5: Hyper-parameters of the VAE . The latent size ds is the one quantity scanned; all five VAE s share every other setting. std.: standardised.
Parameter
Standard
Large
Preprocessing
target
posterior mean zμ , rescaled
scale, noise level
from the training set
Model
architecture
transformer
model dimension
256
512
Appendix
Table 6: Hyper-parameters of the standard and the large LMF ; a value written across both columns holds for both. The network is a transformer acting on the ds latent vectors, with the sparse attention mask of section III.4 . EMA: exponential moving average.
large LMF
ds=64
ds=128
ds=256
ds=512
ds=1024
ds=128
ds=256
ds=512
Geant4
WD/σ\textscGeant4 [ ×10−2 ]
Energy sum
1.2±0.2
1.2±0.3
1.3±0.2
1.5±0.4
2.5±0.7
1.7±0.5
1.1±0.3
2.6±0.5
0.8±0.1
ECal energy sum
1.5±0.5
1.3±0.3
1.3±0.2
1.5±0.5
1.2±0.3
2.1±0.3
1.1±0.3
1.4±0.4
1.0±0.1
HCal energy sum
2.3±0.3
2.2±0.4
1.9±0.2
1.9±0.4
7.5±0.2
1.9±0.3
1.0±0.1
2.5±0.7
1.5±0.6
Number of hits
16.7±0.9
13.7±0.8
15.8±0.8
23.4±0.7
23.5±0.3
13.6±0.6
16.4±0.5
22.0±0.5
1.2±0.3
Centre of gravity x
20.7±0.3
9.3±0.2
14.2±1.1
12.8±0.3
19.9±0.6
8.6±0.9
6.2±1.3
6.5±0.6
1.4±0.3
Appendix
Table 7: Distance between each latent size and Geant4 for every observable, on the full-range samples of 100000 showers: the Wasserstein distance normalised to the Geant4 standard deviation (top) and the Jensen–Shannon divergence (bottom), mean and spread over five disjoint subsets. The best model per observable is in bold, and each block ends with the geometric mean over the observables of the entries of every column, in the units of the block; its error is the standard deviation of that mean over 1000 resamplings of every entry within its uncertainty. The last column is the Geant4 floor.
Figure 9: Linear discriminant analysis of the latent representation, coloured by incident energy, for the posterior mean zμ (left) and the latent z generated by the LMF (right), at ds=64 , 128, 256, 512 and 1024 (top to bottom). The annotation in each panel is the nine-class accuracy of the corresponding classifier, against 1/9≈0.11 for chance.
Figure 10: As figure 9 , for the large LMF at ds=128 , 256 and 512 (left to right).
Figure 11: As figure 5 , for 20000 Geant4 showers (grey) and the same showers compressed and decompressed by the VAE of each latent size, without LayerFM , the LMF and energy calibration. The orange line is the VAE at ds=256 trained without the hit time, absent from the velocity panel. From left to right: mean energy per layer, mean energy per shower in bins of the distance from the shower axis, the energy spectrum of the cells, and the energy per shower in bins of the apparent velocity of the cells.
ds=64
ds=128
ds=256
ds=512
ds=1024
ds=256 , no time
Geant4
WD/σ\textscGeant4 [ ×10−2 ]
Energy sum
2.7±1.3
2.8±1.2
1.9±1.0
2.1±0.7
2.1±0.7
2.1±0.7
2.4±0.9
ECal energy sum
2.4±1.2
1.6±0.5
3.1±0.2
2.9±1.3
2.6±1.0
2.1±0.7
2.8±0.7
HCal energy sum
3.7±1.3
3.1±1.2
2.3±0.7
3.1±1.2
2.2±0.8
3.1±1.2
2.6±0.9
Number of hits
13.2±1.9
11.5±1.2
11.9±0.7
12.9±2.1
12.4±4.0
20.5±1.8
2.2±0.7
Centre of gravity x
20.0±0.7
3.6±0.7
3.6±0.9
3.4±1.4
3.0±0.7
4.7±2.0
3.9±0.7
Centre of gravity y
3.3±1.4
2.8±0.3
2.1±0.3
2.1±0.7
3.0±0.7
2.8±1.4
2.7±1.1
Appendix
Table 8: As table 7 , for 20000 Geant4 showers compressed and decompressed by the VAE of each latent size and by the VAE at ds=256 trained without the hit time, without energy calibration, compared with the same Geant4 showers. The VAE without time has no entry for the three velocities, and its geometric mean, marked † , is over the other eleven observables and not compared with the others. The best latent size per observable is in bold, and each block ends with the geometric mean over the observables of the entries of every column, with its error the standard deviation over 1000 resamplings of the entries within their uncertainties. The last column is Geant4 against itself on subsets of the same size.
Figure 12: Mean number of cells per shower against the time of the first hit in the cell, in logarithmic time bins, for 100000 Geant4 showers (grey), ds=128 (blue) and the large LMF at ds=256 (green), with incident energies between 10 and 90GeV . Top: ECal and HCal together; bottom: ECal (left) and HCal (right). Below each panel, the ratio to Geant4 .
Figure 13: Response, the mean visible energy of the generated showers divided by that of Geant4 , in bins of 2GeV of incident energy. One row per latent size. Left: the LMF , rescaled by a constant factor (points), and the response implied by the fitted factor f(E) (line). Right: the large LMF at ds=128 , 256 and 512, uncalibrated (points), and 1/f(E) of the factor applied to it (line).
Figure 14: As figure 6 , at 20GeV (top) and 50GeV (bottom).
CaloHadronic
AllShowers NFE 32
AllShowers NFE 4
ds=128
ds=256 , large LMF
Geant4
WD/σ\textscGeant4 [ ×10−2 ], at 20 / 50 / 80 GeV
Energy sum
30.3 / 30.7 / 38.8 ± 4.3
11.6 / 9.2 / 7.8 ± 2.3
10.8 / 12.8 / 15.3 ± 3.3
14.8 / 6.8 / 7.8 ± 2.2
11.7 / 8.5 / 6.7 ± 2.8
4.1 / 4.7 / 4.4 ± 2.4
ECal energy sum
7.0 / 6.2 / 7.3 ± 2.3
4.4 / 3.6 / 2.3 ± 1.9
4.0 / 4.2 / 5.0 ± 1.8
4.1 / 3.2 / 3.4 ± 1.2
4.7 / 3.2 / 2.7 ± 1.3
3.1 / 3.7 / 2.4 ± 1.1
HCal energy sum
9.6 / 8.4 / 8.2 ± 3.0
7.2 / 5.9 / 6.0 ± 1.4
7.0 / 6.4 / 8.4 ± 1.7
6.1 / 4.5 / 4.3 ± 1.9
5.7 / 4.3 / 6.2 ± 1.1
2.9 / 2.6 / 4.3 ± 1.9
Number of hits
29.6 / 19.8 / 15.1 ± 2.5
17.1 / 7.5 / 9.8 ± 2.7
28.4 / 18.6 / 17.2 ± 3.2
30.3 / 23.7 / 22.7 ± 3.0
29.8 / 29.7 / 27.3 ± 3.2
3.2 / 3.6 / 3.5 ± 0.9
Centre of gravity x
8.2 / 4.9 / 4.7 ± 2.6
6.7 / 4.0 / 4.0 ± 2.6
12.6 / 13.8 / 9.0 ± 2.1
11.4 / 9.3 / 10.0 ± 3.7
4.2 / 6.8 / 9.6 ± 2.6
5.5 / 5.1 / 5.1 ± 2.7
Appendix
Table 9: Normalised Wasserstein distance and Jensen–Shannon divergence to Geant4 per observable. Each entry gives the values at 20 / 50 / 80GeV with the largest of their three statistical uncertainties, and the best model at each energy is in bold. Each block ends with the geometric mean over the observables, in the units of the block, with the largest of its three errors, each the standard deviation over 1000 resamplings of the entries within their uncertainties, followed by the mean of the three geometric means with the error propagated from theirs. The last column is Geant4 against itself on subsets of the same size.
OOD [%]
Model
AUC
Geant4
generated
AllShowers NFE 32
0.614±0.002
0.13±0.09
0.26±0.12
ds=256 , large LMF
0.702±0.002
0.93±0.31
0.32±0.15
AllShowers NFE 4
0.726±0.003
0.06±0.04
0.20±0.10
CaloHadronic
0.735±0.001
0.13±0.11
2.93±1.05
ds=128
0.755±0.001
3.67±0.31
0.10±0.03
Appendix
Table 10: AUC of the classifier separating each model from Geant4 , and the fraction of Geant4 and of generated test showers classified with more than 99% confidence (OOD); mean and standard deviation over three trainings. Lower is better in every column; the lowest value is in bold.
Figure 15: Linearity (top) and resolution (bottom) of the reconstructed energy of π+ showers for Geant4 , CaloHadronic , AllShowers with 32 and 4 network evaluations (NFE 32, NFE 4), ds=128 and the large LMF at ds=256 , with 10000 showers per energy and generator. Below each, the ratio to Geant4 ; the point beyond the range of the upper ratio panel is marked by a triangle at its edge. At each energy the models are drawn side by side above a bracket.
Hardware
Model
NFE
Batch
Time [s]
Speed-up
CPU
Geant4
–
1
2.09
1.0×
CaloHadronic
1
1
0.59
3.5×
16
0.73
2.8×
59
1
34.8
<0.1×
16
43.3
<0.1×
AllShowers
4
1
0.50
4.2×
Appendix
Table 11: Generation time per shower and speed-up relative to Geant4 on a single CPU core. The Geant4 rows and the CaloHadronic rows at a single network evaluation and on the CPU are taken from ref. [ 18 ] , which takes them from ref. [ 19 ] ; all other rows are measured here as described in the text.
Figure 16: Generation time per shower of the full chain as a function of the number of points per shower (a single network evaluation, batch size 128, NVIDIA A100 GPU), for latent sizes ds from 64 to 1024 and for the large LMF at ds=256 and 512 (“large”).
Figure 17: Generation time per shower of each stage of the chain for the VAE at ds=256 with the standard LMF , and with the large LMF as a dashed line, as a function of the number of points per shower (a single network evaluation, batch size 128, NVIDIA A100 GPU). The dotted line is the VAE decompressor without its two refinement blocks, the self-attention blocks at full resolution that close it, labelled “last layers” in the legend.
Figure 18: Two π+ showers in three dimensions, Geant4 (left), its reconstruction by the VAE (centre) and the shower generated from the same conditioning (right): (a) a shower with a second, later cluster higher in the HCal, connected to the core by tracks, and (b) a shower with secondary tracks leaving the core at large angles.
Figure 19: As figure 18 : (c) a shower whose secondary tracks leave the core towards the back of the HCal, (d) a primary track that crosses almost the whole calorimeter before the shower starts, and (e) a shower in the HCal followed by a long secondary track and a second interaction at the back.
Accurate and efficient detector simulation is essential for modern collider experiments. To reduce the high computational cost, various fast machine learning surrogate models have been proposed. Traditional surrogate models for calorimeter shower modeling train separate networks for each particle species, limiting scalability and reuse. We introduce AllShowers, a unified generative model that simulates calorimeter showers across multiple particle types using a single generative model. AllShowers is a continuous normalizing flow model with a Transformer architecture, enabling it to generate complex spatial and energy correlations in variable-length point cloud representations of showers. Trained on a diverse dataset of simulated showers in the highly granular ILD detector, the model demonstrates the ability to generate realistic showers for electrons, photons, and charged and neutral hadrons across a wide range of incident energies and angles without retraining. In addition to unifying shower generation for multiple particle types, AllShowers surpasses the fidelity of previous single-particle-type models for hadronic showers. Key innovations include the use of a layer embedding, allowing the model to learn all relevant calorimeter layer properties; a custom attention masking scheme to reduce computational demands and introduce a helpful inductive bias; and a shower- and layer-wise optimal transport mapping to improve training convergence and sample quality. AllShowers marks a significant step towards a universal model for calorimeter shower simulations in collider experiments.
Thorsten Buss, Henry Day-Hall, Frank Gaede +2
Institut für Experimentalphysik, Universität Hamburg, Hamburg, Germany · Deutsches Elektronen-Synchrotron DESY, Hamburg, Germany
High-precision calorimeter simulation at current and future colliders imposes rapidly growing computational demands, motivating the development of machine-learning surrogates for traditional Monte Carlo tools such as Geant4. Flow matching and diffusion-based generative models have become leading approaches for high-dimensional fast simulation because of their sample quality, but typically require O(100) function evaluations at inference and often rely on auxiliary networks to constrain global observables, compromising streamlined end-to-end generation. We introduce a unified framework that improves the balance between speed, shower quality, and physics fidelity. The method combines: (i) an average velocity field integrator that enables sampling in one or a few evaluations; (ii) a learned generative prior in shower space, constructed from data rather than random noise; and (iii) physics-guided loss terms that impose inductive biases on key observables during training. These elements are training time regularizers, preserving end-to-end inference with no additional cost. With only one or a few evaluation steps, the model achieves shower quality competitive with state-of-the-art flow and diffusion approaches, tested on several public high granularity calorimeter datasets. The results demonstrate inter-layer shower structure consistent with the underlying physics, providing a strong candidate for future fast simulation workflows.
Cheng Jiang, Sitian Qian, Kevin Pedro +3
School of Physics and Astronomy, University of Edinburgh, Edinburgh EH9 3FD, United Kingdom · Department of Physics, University of Wisconsin-Madison, Madison, WI 53706, USA · Fermi National Accelerator Laboratory, Batavia, IL 60510, USA +2
In High Energy Physics, detailed calorimeter simulations and reconstructions are essential for accurate energy measurements and particle identification, but their high granularity makes them computationally expensive. Developing data-driven techniques capable of recovering fine-grained information from coarser readouts, a task known as calorimeter superresolution, offers a promising way to reduce both computational and hardware costs while preserving detector performance. This thesis investigates whether a generative model originally designed for fast simulation can be effectively applied to calorimeter superresolution. Specifically, the model proposed in arXiv:2308.11700 is re-implemented independently and trained on the CaloChallenge 2022 dataset based on the Geant4 Par04 calorimeter geometry. Finally, the model's performance is assessed through a rigorous statistical evaluation framework, following the methodology introduced in arXiv:2409.16336, to quantitatively test its ability to reproduce the reference distributions.
Andrea Cosso
Università degli Studi di Genova Scuola di Scienze Matematiche, Fisiche e Naturali Laurea Magistrale in Fisica