A unified surrogate model for enhanced photon shielding through an all-natural-element multi-group flux dataset

Efficient gamma shielding analysis is vital for nuclear safety. However, traditional Point Kernel approaches relying on outdated ANSI databases face significant limitations in providing comprehensive and fine-grained shielding analysis. This study addresses this by creating a detailed multi-group photon flux dataset for 92 nuclides via Monte Carlo simulations. The dataset’s complexity renders traditional modeling techniques ineffective. We introduce a novel generative-reconstruction surrogate model, combining a conditional Generative Adversarial Network (cGAN) and a fine-tuned UNet, both enhanced with self-attention mechanisms. This model predicts complex multi-group photon shielding parameter fields. Verification shows the model accurately predicts parameter fields, with 95% of samples achieving an average relative deviation below 20%. Predicted relative flux, converted to buildup factors, aligns well with Monte Carlo truth and ANSI values, confirming reliability and improved conservatism. This approach offers an efficient, accurate alternative for photon shielding calculations, proposing a new approach for data and computation.