A rapid atmospheric radionuclide dispersion prediction and dose assessment method based on the physics-aware spatio-temporal neural operator.

Abstract

Rapid prediction of atmospheric radionuclide dispersion is essential for emergency decision-making and radiation protection in nuclear accidents. Although computational fluid dynamics (CFD) methods provide high accuracy, their limited efficiency makes it difficult to meet timeliness requirements. To address this limitation, this study proposed a method for rapid atmospheric radionuclide dispersion prediction and radiation dose assessment based on a physics-aware spatio-temporal neural operator (PA-STNO). Within the proposed model-based method, condition encoding represents variations in meteorological and release conditions, spatio-temporal attention enhances plume-evolution and spatial-structure features, the Fourier Neural Operator backbone models long-range transport dependencies, and a physics-aware composite loss improves the representation of high-concentration regions, plume boundaries, and overall concentration-mass consistency. Multi-condition tritium dispersion data generated with OpenFOAM were used for model training and evaluation. On the primary test set, PA-STNO achieved a Relative L2 of 0.21, an FAC2 of 0.92, and a Plume IoU of 0.82, while substantially improving prediction speed. The model also outperformed the baseline models, and a supplementary comparison with a conventional Lagrangian dispersion model showed reasonable agreement, with a mean absolute relative deviation of 8.36%. The predicted fields were further used for radiation dose assessment, linking rapid concentration prediction with dose calculation and supporting the preliminary assessment of dispersion consequences and radiological impacts.