Physics-Informed RecurrentNet for Fast and High-Fidelity Atmospheric Dispersion Simulation of Multi-Nuclide in Nuclear Emergencies

摘要: Rapid and high-fidelity forecasting of atmospheric radioactive nuclide dispersion is critical for nuclear emergency responses. Traditional physics-based models struggle with an inherent trade-off between fidelity and computational speed. While data-driven deep learning (DL) offers a fast surrogate alternative, generic spatio-temporal DL architectures suffer from severe "topological collapse" (e.g., over-smoothing and blurred boundaries) when confronted with highly sparse (>80% zero-value background) and heavy-tailed (spanning 20 orders of magnitude) radioactive plume fields. To bridge this gap, we propose a novel Physics-Informed RecurrentNet (PIRNet). Specifically, we design a dual spatial physical prior mechanism-integrating a source-injected Gaussian heatmap with absolute coordinate grids-to explicitly break spatial symmetry and anchor the physical diffusion origin. Furthermore, we develop an end-to-end architecture driven by ConvGRU and a PixelShuffle decoder to implicitly model the Markovian advection-diffusion process, effectively suppressing checkerboard artifacts. Finally, we introduce a composite physics-constrained loss featuring a spatial gradient error (SGE) penalty to reconstruct sharp fluid-dynamic fronts. Comprehensive experiments demonstrate that PIRNet provides a favorable balance between prediction accuracy and computational efficiency. With a highly lightweight profile of merely 2.78 M parameters, PIRNet accelerates inference by approximately 900 times compared to the high-order Lagrangian Particle Dispersion Model (LPDM), requiring only 0.207 seconds per case. Moreover, it consistently achieves better performance than the selected baseline models (Flattened-LSTM, 3D-CNN, STGCN) in spatial morphological fidelity. It uniquely suppresses the SGE below 0.1 (0.0962) and achieves a remarkable Plume Factor of 10 (FAC10) accuracy of 62.76%. This study provides a potential digital-twin-based framework for multi-nuclide dispersion forecasting in nuclear safety management.