AllShowers-multi-geometry

AllShowers, multi-geometry pre-trained models

Transformer-based flow-matching generative models for electromagnetic calorimeter showers, represented as point clouds. These are the two multi-geometry pre-trained models from the paper Transferable Fast Calorimeter Shower Generation via Multi-Geometry Pre-training (arXiv:2608.18233): they are the starting points that the paper fine-tunes to new detector geometries with a fraction of the data and compute of training from scratch.

The model generates the shower point cloud conditioned on the incident particle energy, the direction (unit vector), the calorimeter sampling fraction, and the number of layers. Per-layer point counts are provided by the companion count model FLC-QU-hep/PointCountFM-multi-geometry.

Checkpoints

Both share the same architecture (transformer flow matching, embedding 64, 4 heads, 4 blocks, feedforward 256) and conditioning. They differ in the maximum number of layers and in the pre-training data. See the paper for the comparison between the two pre-training strategies.

Files and usage

Each folder is a self-contained run directory in the layout expected by the AllShowers code:

<folder>/
├── conf.yaml               # architecture + transform definitions
├── weights/best.pt         # best-validation model state_dict
└── preprocessing/trafos.pt # fitted input transformations (required!)

Download a folder and pass it as the run directory to the generation entry point of the AllShowers repository:

from huggingface_hub import snapshot_download
run_dir = snapshot_download("FLC-QU-hep/AllShowers-multi-geometry",
                            allow_patterns="lemurs/*") + "/lemurs"
# then follow the generation instructions in the AllShowers repository,
# passing run_dir as the model/run directory

Note: preprocessing/trafos.pt contains the fitted normalization of all inputs. The weights are unusable without it.

Training data

The pre-training datasets (Geant4, LEMURS + SimpleBox) are published at doi:10.25592/uhhfdm.19103.

Citation

If you use these weights, please cite:

@article{Buss2026b,
    author = {Buss, Thorsten and Day-Hall, Henry and Gaede, Frank and Kasieczka, Gregor and Kr{\"u}ger, Katja and McKeown, Peter and Valente, Lorenzo},
    title = "{Transferable Fast Calorimeter Shower Generation via Multi-Geometry Pre-training}",
    eprint = "2608.18233",
    archivePrefix = "arXiv",
    primaryClass = "physics.ins-det",
    doi = "10.1088/1748-0221/21/10/P10041",
    journal = "JINST",
    volume = "21",
    number = "10",
    pages = "P10041",
    month = "10",
    year = "2026"
}