2026年“核理安邦”联合博士生学术论坛

一种用于三维辐射场精细重建的深度学习框架

Not scheduled

12m

地址：清华大学校内

口头报告 工程博士论坛 工程博士论坛

Speaker

豪 胡

摘要

核能是全球清洁能源组合的关键组成部分。在核设施的设计、运行、退役及后处理等全生命周期中，准确表征三维空间剂量率分布对于保障作业人员辐射安全及优化屏蔽方案至关重要。然而，受限于探测成本与环境限制，实际工程中的测点通常极度稀疏，使得获取高分辨率辐射场数据面临巨大挑战。本研究提出了一种通用的深度学习框架 PatchUNet3D，旨在通过极少量测点实现高精度三维辐射场重建。该框架的核心创新在于 Patchify Stem 模块。该模块采用单步大步长（stride=5）三维卷积，替代了标准 UNet3D 前两级的全分辨率编码操作，将极稀疏输入的有效信息密度提升了约 125 倍，从根本上解决了标准架构在处理海量零值体素时参数更新效率低下的瓶颈。在参数量相近（约 5.8M）的条件下，本研究建立了 PatchUNet3D 与标准 UNet3D 的严格消融对比。以核电站放射性废物桶暂存库为典型验证场景，在复杂度递增的蒙特卡洛仿真数据集上进行的系统评估结果表明，所提框架在重建精度、高剂量区域保真度及鲁棒性方面均显著优于基准模型。实验证实，PatchUNet3D 在仅有 1 到 10 个测点的条件下仍能保持优异的性能，大幅降低了对密集传感器网络的依赖。本研究为核设施低成本、实时辐射监测提供了新的技术范式，在提升核安全保障与运维效率方面具有显著的应用价值

Abstract

Nuclear energy is a critical component of the global clean energy portfolio. The safety and efficiency of its lifecycle, particularly during radioactive waste disposal, rely on the accurate characterization of the three-dimensional (3D) radiation field. However, the sparse distribution of measurements often makes it difficult to obtain high-resolution radiation field data. This study addresses this challenge by developing an innovative PatchUNet3D deep learning framework for high-accuracy 3D radiation field reconstruction from sparse measurements. The core innovation of the proposed method is the Patchify Stem module — a single large-stride (stride=5) 3D convolution that replaces the first two full-resolution encoding stages of the standard UNet3D, boosting the effective information density of the sparse input by a factor of ~125×. The effectiveness of the proposed method is evaluated by comparing its performance against a conventional single-step UNet3D architecture designed for the same end-to-end prediction task. Our results demonstrate that, with nearly identical parameter counts, PatchUNet3D significantly outperforms the UNet3D baseline. Moreover, this method significantly reduces the reliance on dense sensor networks, enabling more cost-effective monitoring systems. The framework's performance was rigorously evaluated on a simulated dataset representing diverse scenarios within a nuclear waste disposal facility. The results confirm the model's high accuracy and robustness, demonstrating its substantial application value in advancing the safety and efficiency of nuclear waste disposal operations.