Cross-Geometry Transfer Learning in Fast Electromagnetic Shower Simulation

TL;DR

Calorimeter shower simulation is a major computational bottleneck. The authors propose cross-geometry transfer learning using diffusion-based, point-cloud shower generation, pre-trained on ILD data, to adapt to diverse detector geometries with limited target data. They demonstrate substantial improvements in distribution fidelity (notably a ~44% reduction in the geometric mean Wasserstein distance) for CaloChallenge in low-data regimes and show BitFit as a practical, parameter-efficient fine-tuning approach, while LoRA underperforms due to high intrinsic dimensionality in physics transformations. This work provides a step toward a fast, geometry-robust FastSim foundation for calorimeter simulations and highlights both potential and limits of current PEFT strategies in physics-based generative modeling.

Abstract

Accurate particle shower simulation remains a critical computational bottleneck for high-energy physics. Traditional Monte Carlo methods, such as Geant4, are computationally prohibitive, while existing machine learning surrogates are tied to specific detector geometries and require complete retraining for each design change or alternative detector. We present a transfer learning framework for generative calorimeter simulation models that enables adaptation across diverse geometries with high data efficiency. Using point cloud representations and pre-training on the International Large Detector detector, our approach handles new configurations without re-voxelizing showers for each geometry. On the CaloChallenge dataset, transfer learning with only 100 target-domain samples achieves a $44\%$ improvement on the geometric mean of Wasserstein distance over training from scratch. Parameter-efficient fine-tuning with bias-only adaptation achieves competitive performance while updating only $17\%$ of model parameters. Our analysis provides insight into adaptation mechanisms for particle shower development, establishing a baseline for future progress of point cloud approaches in calorimeter simulation.

Figures (17)

• Figure 1: The transfer learning approach presented in this work. A model pre-trained on the ILD detector is adapted to new geometries, such as CaloChallenge Dataset 3, through fine-tuning. This approach contrasts with the conventional "from scratch" paradigm, where models are initialised with random weights and must learn all physics representations directly from the target dataset. The dashed box with a question mark represents potential future applications to additional detector configurations.
• Figure 2: Incident energy distributions for pre-training (ILD, red, uniform 10-90 GeV) and downstream (CaloChallenge, blue, log-uniform 1-1000 GeV) datasets. Left: Full range, with a dashed box indicating the overlap region. Right: Magnified overlap showing distributional differences that, combined with particle type and geometry shifts, constitute the compound domain shift addressed in this work.
• Figure 3: Representative electromagnetic shower event displays illustrating the domain shift. Left: 81 GeV photon shower in the planar ILD detector. Right: 913 GeV electron shower in the cylindrical CaloChallenge detector. The cylindrical layer structure is visible in the curved distribution of energy deposits along the longitudinal axis. Data representation from Ref. Buss:2024orz.
• Figure 4: ShowerFlow transfer performance measured by normalised Wasserstein distance between generated and reference point-count distributions, averaged across all 45 calorimeter layers. Each point represents the median performance across five independent training runs with different random seeds. Error bands show the standard deviation across seeds. Evaluation is performed on the full 10,000-sample validation set. Fine-Tuning from ILD-pretrained weights substantially outperforms training From Scratch in low-data regimes.
• Figure 5: Geant4 vs generated showers at training sizes $D$. Top rows: from scratch; bottom rows: full fine-tuned. All histograms from $10,000$ events with energy logarithmically distributed from $1-1000$ GeV. Bottom panels show Geant4 ratios. The error band corresponds to the statistical uncertainty in each bin.