3D Radiation Field Reconstruction Method Based on U-Net

Digital transformation of radiation protection management in nuclear facilities relies on accurate 3D radiation field reconstruction techniques. While traditional reconstruction methods face significant limitations when dealing with sparse measurement data, neural networks have emerged as a promising alternative. However, fully-connected neural networks (FCNN) possess an inherent architectural flaw for this task: They require the flattening of 3D spatial data into a one-dimensional vector, which destroys the intrinsic spatial correlations crucial for accurate field representation. To address this critical gap, a sparse input-dense output 3D U-Net architecture—designed specifically to preserve and leverage these spatial relationships through its convolutional encoder-decoder structure and skip-connections—was proposed and evaluated in this research. To validate the proposed method, a rigorous comparative study between the U-Net model and a conventional FCNN was designed. The methodology involved the creation of extensive datasets for two distinct scenarios: a simple case featuring a single radiation source and a single shielding wall, and a complex case designed to mimic a realistic nuclear facility room with multiple radiation sources and intricate shielding structures. For these scenarios, a total of 50 000 data samples for each were generated using a custom-developed point kernel integration program, CIRPDose. The simulated space was discretized into a grid of 11×11×11 (1 331) points. Sparse input data was generated by randomly selecting a small fraction of these points to act as virtual detectors, with a sampling rate of approximately 2% (27 points) for the simple scenario and 3% (42 points) for the complex one. The results of the comparative analysis demonstrate the definitive superiority of the U-Net architecture. In the simple ‘single-source, single-shielding’ scenario, the U-Net model achieves a mean absolute percentage error (MAPE) of 4.48%, a notable improvement over the FCNN’s MAPE of 6.74%. This performance advantage becomes substantially more pronounced in the ‘multi-source, multi-shielding’ complex scenario. Here, the U-Net maintains a high level of accuracy with a MAPE of 7.50%, whereas the FCNN’s performance degrades significantly, resulting in a MAPE of 21.99%. Examination of the training curves also reveals that while the FCNN is prone to overfitting in the complex scenario, the U-Net exhibits better generalization capabilities. In conclusion, this research successfully demonstrates that the proposed sparse input-dense output U-Net model is a highly effective and robust method for 3D radiation field reconstruction, significantly outperforming the standard FCNN approach, particularly in complex environments with multiple sources and shields. By enabling accurate field reconstruction from very sparse data, this method provides critical algorithmic support for the development of advanced, digitally-driven radiation protection systems in nuclear facilities. While the model shows high overall accuracy, it should be noted that reconstruction errors remain a concern in regions with high dose gradients, suggesting that future work could explore strategies such as adaptive mesh refinement to further enhance precision in these critical areas.