Simulated MeV-UED diffraction image dataset for machine-learning-based multidimensional electron beam characterization

This dataset contains the simulated ultrafast electron diffraction (UED) image data and the accompanying code used in the study "High-dimensional electron beam characterization via machine learning of diffraction images." The data support a ResNet-50-based deep-learning framework that reconstructs multiple electron-beam parameters directly from single-shot diffraction patterns of a polycrystalline aluminum reference foil. The dataset comprises two parts, generated from beam-dynamics simulations (ASTRA) combined with a kinematical polycrystalline scattering model, based on the MeV-UED platform at Huazhong University of Science and Technology: 1.

Transverse set (horizon): 38,400 single-frame diffraction images from 400 beam configurations, with per-image labels for horizontal/vertical beam size, horizontal/vertical angular divergence, and energy spread. Parameter ranges: beam size 0.014-1.28 mm, angular divergence 0.005-3.34 mrad, energy spread 2.39-23.08 keV. 2. Longitudinal set: 3,555 time-resolved diffraction images, with per-image labels (train.csv + test.csv) for bunch length over the range 96-495 fs, and a per-bunch master table (labels.csv). Also included are the data-generation scripts, ASTRA input templates, and the training/evaluation code (Python, with English comments) needed to reproduce the results, organized so that each network finds its data under dataset/.