LayerFM
Authors: Martina Mozzanica, Thorsten Buss
LayerFM generates the per-layer structure of a hadronic shower: for each of the 78 layers of the ILD calorimeter, the number of points (hits) and, for the time model, the earliest hit time in that layer, conditioned on the incident energy. It is the first stage of a fast point-cloud shower generator, where these per-layer quantities fix how many points the later stages produce.
Each model is a conditional flow-matching network (CNF). The recommended files are its distilled one-step student, which generates in a single network evaluation.
Code, training and distillation: github.com/FLC-QU-hep/LayerFM
Models
| Folder | Output per layer | Input size | Parameters |
|---|---|---|---|
20260619_124529_LayerFM_PandT_/ |
number of points + minimum hit time | 156 | 0.75 M per network |
20260619_192637_LayerFM_P/ |
number of points | 78 | 0.30 M per network |
Each folder contains:
| File | Content |
|---|---|
compiled_distill.pt |
distilled one-step student, TorchScript (recommended) |
compiled.pt |
CNF teacher, TorchScript, 50 Heun steps |
checkpoint_distill.pt, checkpoint.pt |
student and teacher state dicts, with optimizer state |
conf.yaml |
training configuration |
The compiled*.pt files are self-contained: the fitted preprocessing is built in, so they
run with PyTorch alone, without the LayerFM code or the training data. The checkpoints are
for fine-tuning or further training with the LayerFM code; rebuilding their preprocessing
needs a training file (set data.data_file in conf.yaml).
Usage
import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download("FLC-QU-hep/LayerFM", "20260619_124529_LayerFM_PandT_/compiled_distill.pt")
model = torch.jit.load(path, map_location="cpu")
energy = torch.full((1000, 1), 50.0) # incident energy in GeV, shape (N, 1)
with torch.no_grad():
num_points, min_time = model(energy) # each (N, 78)
num_points = num_points.clamp(min=0).round().int()
20260619_192637_LayerFM_P/ models return a single tensor of shape (N, 78). Minimum times are in ns;
layers without points have no meaningful time.
Training data
Geant4 11.2.2 (QGSP_BERT) showers of single π⁺ with energies uniform between 10 and
90 GeV, entering the ILD calorimeter perpendicularly: a silicon–tungsten ECal of 30 layers
(5 × 5 mm² cells) and a scintillator–steel HCal of 48 layers (3 × 3 cm² cells). The
models were trained on 1 806 445 showers, with 50 000 held out for validation, for 1500
epochs. The student was then trained for 1500 epochs to reproduce the teacher's
200-step Heun samples from the same noise.
Scope: the models are only valid for this detector, single π⁺ at normal incidence and 10–90 GeV.
Citation
LayerFM builds on CaloHadronic (arXiv:2506.21720, JINST 21 (2026) P01042) and on PointCountFM, the point-count model of AllShowers. If you use these models, please cite both:
@article{Buss:2025cyw,
author = {Buss, Thorsten and Gaede, Frank and Kasieczka, Gregor and Korol, Anatolii and Kr{\"u}ger, Katja and McKeown, Peter and Mozzanica, Martina},
title = {{CaloHadronic: a diffusion model for the generation of hadronic showers}},
eprint = {2506.21720},
archivePrefix = {arXiv},
primaryClass = {physics.ins-det},
doi = {10.1088/1748-0221/21/01/P01042},
journal = {JINST},
volume = {21},
number = {01},
pages = {P01042},
year = {2026}
}
@article{Buss:2026yrf,
author = {Buss, Thorsten and Day-Hall, Henry and Gaede, Frank and Kasieczka, Gregor and Kr{\"u}ger, Katja},
title = {{AllShowers: One model for all calorimeter showers}},
eprint = {2601.11716},
archivePrefix = {arXiv},
primaryClass = {physics.ins-det},
journal = {arXiv e-prints},
month = {1},
year = {2026}
}
License
MIT