Calorimeter Shower Dataset and Diffusion Model Checkpoints for Gradient-Based Detector Optimization

This dataset accompanies the paper “Differentiable Surrogate for Detector Simulation and Design with Diffusion Models”, available at Machine Learning: Science and Technology (2026), DOI: https://doi.org/10.1088/2632-2153/ae5c56. This dataset contains preprocessed photon shower simulation data generated from Geant4-based calorimeter simulations, used for training and evaluating conditional diffusion models in gradient-based detector optimization. Each event represents the energy deposition profile of an electromagnetic shower under varying calorimeter configurations. The raw ROOT files were converted into normalized 2D histograms using Python with the uproot, numpy, matplotlib, and h5py libraries. For each event, two histograms represent the energy-deposition maps in orthogonal planes: images_xz: Energy deposition in the x–z plane. images_yz: Energy deposition in the y–z plane. Each histogram contains 100×100 bins, normalized per event. Labels and conditioning variables: labels_energy: Incident photon energies mapped as {1, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 GeV}. labels_xy: Transverse cell size (granularity) of the calorimeter: {1, 2, 3, 4, 5 cm}. labels_z: Longitudinal cell depth: {4, 5, 8, 10, 15 cm}. labels_material: Absorber material type (PbF₂). File organization: total_photon_shower: Dataset used for diffusion model pre-training (100k events). post_total_photon_shower: Reduced dataset for LoRA fine-tuning (10k events). Generation details: Binning: 100×100 per 2D histogram. Automatic range scaling depending on detector geometry. Purpose:Designed for developing and benchmarking conditional diffusion models for calorimeter shower generation, surrogate modeling, and differentiable optimization of detector design parameters (e.g. geometry, segmentation, material). Format:HDF5 (.h5), fully compatible with Python’s h5py, numpy, and pandas.本数据集配套发表于《Machine Learning: Science and Technology》2026年的论文《面向探测器模拟与设计的扩散模型可微代理》（Differentiable Surrogate for Detector Simulation and Design with Diffusion Models），DOI：https://doi.org/10.1088/2632-2153/ae5c56。 本数据集包含基于Geant4的量能器（calorimeter）模拟生成的预处理光子簇射模拟数据，用于基于梯度的探测器优化中条件扩散模型的训练与评估。每个事件代表不同量能器配置下的电磁簇射能量沉积分布。 原始ROOT文件通过Python结合uproot、numpy、matplotlib与h5py库转换为归一化二维直方图。每个事件对应正交平面上的两张能量沉积图： images_xz：x-z平面内的能量沉积分布 images_yz：y-z平面内的能量沉积分布 每张直方图包含100×100个分箱，且按事件进行归一化。 标签与条件变量： labels_energy：入射光子能量，取值为{1、10、20、30、40、50、60、70、80、90、100 GeV} labels_xy：量能器横向单元尺寸（粒度），取值为{1、2、3、4、5 cm} labels_z：纵向单元深度，取值为{4、5、8、10、15 cm} labels_material：吸收体材料类型（PbF₂） 文件组织结构： total_photon_shower：用于扩散模型预训练的数据集，包含10万条事件 post_total_photon_shower：用于LoRA（Low-Rank Adaptation）微调的精简数据集，包含1万条事件 生成细节： 分箱设置：每张二维直方图为100×100分箱 根据探测器几何结构自动进行范围缩放 应用场景：本数据集旨在开发与基准测试评估用于量能器簇射生成、代理建模以及探测器设计参数（如几何结构、分割方式、材料）可微优化的条件扩散模型。 数据格式：HDF5（.h5）格式，可完全兼容Python的h5py、numpy与pandas库。