Multi-Geometry Pre-training Cuts Calorimeter Simulation Error 5x on Unseen Detectors

The arXiv paper demonstrates that geometric diversity during pre-training produces transferable surrogates for calorimeter shower simulation. Researchers generated SimpleBox, a family of 10,000 box detectors spanning sampling fraction and longitudinal segmentation, then compared it against pre-training on realistic detector layouts. Both priors were evaluated on an unseen target calorimeter using fine-tuning sets from 10^3 to 10^5 showers. The synthetic prior outperformed training from scratch and, at larger fine-tuning sizes, surpassed the realistic-detector prior, showing that explicit variation in sampling and segmentation teaches more generalizable shower physics than exposure to any single real geometry.

High-energy physics experiments currently spend the majority of their computing budget on Geant4 calorimeter simulations. Generative surrogates have historically been detector-specific, requiring fresh large datasets for each upgrade or new experiment. By decoupling the pre-training distribution from any fixed detector, the work lowers the data threshold for deploying fast simulation. This directly addresses the projected computing shortfall for HL-LHC and future colliders where shower libraries must be regenerated for evolving calorimeter designs.

The study leaves open whether the observed transfer gains persist when the target geometry differs in material composition or readout segmentation beyond the SimpleBox plane. A follow-up experiment that includes active-layer thickness and absorber density as additional pre-training axes would test whether further geometric axes compound the benefit. Integration into production frameworks such as CMSSW or ATLAS FastCaloSim will also require validation of physics observables beyond sliced Wasserstein distance.