2026-10-14T18:22:55Z https://www.scidb.cn/oai

10.57760/sciencedb.j00186.000zf 2026-10-10T13:01:06Z

2026-10-10 Data and code for residual U-Net denoising of three-dimensional radiation fields from low-particle-count Monte Carlo simulations doi:10.57760/sciencedb.j00186.000zf en This dataset provides the data and code supporting the study entitled “Residual U-Net denoising for refining three-dimensional radiation fields from low-particle-count Monte Carlo simulations.” It supports the development and evaluation of residual U-Net denoising to refine fields generated by low-particle-count Monte Carlo simulations and approximate the corresponding high-particle-count reference fields. Denoising is performed on the original spatial grid without changing its resolution.The dataset contains two simulation cases: photon transport in a waste-drum storage room and neutron transport in a room with nested shielding structures. For each case, 100 configurations were generated through random parameter sampling. Monte Carlo simulations with 10⁶ and 10⁹ particle histories were performed for each configuration to produce paired noisy inputs and reference fields, respectively. The 100 pairs in each case were randomly divided into 70 training, 15 validation, and 15 test pairs.For the photon case, dose-rate fields were tallied on a 200 × 50 × 70 grid with a voxel size of 5 cm. The varied parameters include the number and positions of the waste drums, together with the density, activity, and photon energy spectrum of the source material within them. For the neutron case, dose-rate fields were tallied on a 200 × 200 × 100 grid with a voxel size of 5 cm. The room contains an inner cylindrical shield and an outer shield with a square cross-section, with one point source in each of three distinct air regions. The varied parameters include source positions and energies, as well as shielding dimensions, thicknesses, and materials.The deposited files include Monte Carlo simulation inputs and configuration parameters, three-dimensional radiation-field data, preprocessing scripts, and code for training, inference, evaluation, and visualization using Res-UNet, U-Net, and RED-CNN. Preprocessing includes handling nonpositive values, natural-logarithm transformation, and normalization. The dataset supports reproduction of the model training, model comparisons, and quantitative evaluation reported in the study, as well as further research on three-dimensional radiation-field denoising. The principal file formats are TXT, NPY, Python scripts (.py), JSON, and Markdown Monte-Carlo; deep learning; radiation fields; Denoising; U-Net Hu Hao Xu Yuhan Li Xinpeng Luo Haoyuan Fang Sheng PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank