Reconstruction of Power Distribution in Compact Reactors Using EX-Vessel Detection Via Convolutional Neural Networks and Generative Adversarial Network Framework

For compact reactors, the inability to deploy neutron detectors internally due to structural volume constraints makes reconstructing core power distribution using ex-vessel detector data critical for safe operation. In this study, a model of the reactor was developed using OpenMC software. Back-propagation (BP) neural network was employed to reconstruct radial and axial power distributions based on ex-vessel detector signals. By implementing an axially segmented detector method, the average relative error in axial power distribution was reduced from 19.29% to 3.62%. Further integration of convolutional layers and an attention mechanism module decreased the overall average relative error to 0.78%. A two-stage optimized neural network framework was subsequently applied to reconstruct the three-dimensional core power distribution layer-by-layer, achieving error of 1.13%, which meets the requirements for power distribution reconstruction in compact reactors. Notably, this work pioneers the incorporation of a generative adversarial network (GAN) mechanism into power distribution reconstruction, demonstrating the effectiveness of high-quality virtual samples in suppressing errors under specific operational conditions. This GAN-enhanced framework establishes a physics-informed methodology to enhance reconstruction accuracy under data-scarce reactor conditions, particularly for compact cores with ex-vessel-only monitoring.