August 17, 2026 Dataset Open Access

Point cloud calorimeter shower datasets for multi-geometry pre-training: SimpleBox and LEMURS

Lorenzo Valente; Thorsten Buss; Henry Day-Hall; Frank Gaede; Gregor Kasieczka; Katja Krüger; Peter McKeown

Contact person(s)

This record contains the simulated calorimeter shower datasets used in the study of 'Transferable Fast Calorimeter Shower Generation via Multi-Geometry Pre-training'.

All showers are photon-induced electromagnetic showers with incident energy uniformly distributed between 1 and 100 GeV, stored as HDF5 point clouds: each point carries the two transverse displacements with respect to the impact point, the longitudinal layer index, and the deposited energy. Deposits are clustered on a 1 mm × 1 mm transverse grid per layer, and cells below 10 keV are discarded. All detectors are described in DD4hep and simulated with Geant4 11.2.2 through the ddsim driver within the Key4hep stack.

• SimpleBox is a synthetic family of ten thousand box-shaped tungsten-silicon sampling calorimeters spanning sampling fraction, longitudinal segmentation and incident angle, with four million showers in total, plus a one hundred thousand shower subsample, a test set of eight held-out configurations, and a held-out in-range configuration for zero-shot evaluation.
• LEMURS is a point cloud re-simulation of five realistic barrel electromagnetic calorimeters (FCCee-CLD, the Open Data Detector calorimeter, Par04-SiW, Par04-SciPb and the noble-liquid FCCee-ALLEGRO), extracted on a fine transverse grid rather than the cell-aggregated representation of the original release. The first four provide one million training showers each and form the realistic pre-training pool. FCCee-ALLEGRO is the transfer target, with a one hundred thousand shower fine-tuning dataset. Every geometry also has an independent ten thousand shower test set, simulated with disjoint seeds.

The structure is encoded in the file names. Per-file sizes, shower counts and MD5 checksums are in the included MANIFEST.json, and the full schema is documented in README.md.