Omar Ahmed, Elhadad Mohamed K, El-Samrah Moamen G, Nagla Tarek F, Mekkawy Tamer
Nuclear Engineering Department, Military Technical College, Kobry El-Kobbah, Cairo, Egypt.
Department of Computer Engineering and AI, Military Technical College, Kobry El-Kobbah, Cairo, Egypt.
Sci Rep. 2025 Aug 18;15(1):30173. doi: 10.1038/s41598-025-13794-7.
This paper presents a novel approach, ReactorNet, a machine learning framework leveraging thermal neutron flux imaging to enable real-time monitoring of pressurized water reactors (PWRs). By integrating EfficientNetB0 with a hybrid classification-regression architecture, the model accurately identifies control rod positions and operational parameters through thermal neutron flux patterns detected by ex-core sensors. Principal Component Analysis (PCA) and Clustering Analysis decode radial flux variations linked to rod movements, while simulations of a 2772-MW(th) PWR using TRITON FORTRAN validate the framework. This framework outperforms Vision Transformers and ResNet50, achieving superior multi-class accuracy (97.5%) and reduced the mean absolute error (MAE) of regression. Test-Time Augmentation and cross-validation mitigate data limitations, ensuring robustness. This work bridges AI and nuclear engineering, demonstrating EfficientNetB0's potential for precise, real-time reactor monitoring, enhancing operational safety and efficiency.
本文提出了一种新颖的方法——反应堆网络（ReactorNet），这是一种利用热中子通量成像的机器学习框架，用于对压水反应堆（PWR）进行实时监测。通过将高效网络B0（EfficientNetB0）与混合分类回归架构相结合，该模型通过堆外传感器检测到的热中子通量模式，准确识别控制棒位置和运行参数。主成分分析（PCA）和聚类分析对与棒运动相关的径向通量变化进行解码，而使用TRITON FORTRAN对一个2772兆瓦（热）的压水反应堆进行的模拟验证了该框架。该框架优于视觉Transformer和ResNet50，实现了卓越的多类准确率（97.5%），并降低了回归的平均绝对误差（MAE）。测试时增强和交叉验证减轻了数据限制，确保了稳健性。这项工作架起了人工智能与核工程之间的桥梁，展示了高效网络B0在精确、实时反应堆监测方面的潜力，提高了运行安全性和效率。
