Physics-informed deep learning for rapid marine radionuclide dispersion forecasting in nuclear emergency response.

In the event of a severe nuclear accident at a coastal nuclear power plant, the rapid and accurate assessment of radionuclide dispersion in surrounding coastal waters is critical for effective emergency response. To overcome the inherent latency and predictive fidelity limitations of conventional marine dispersion simulations, this study develops an innovative hybrid Physics-Informed Deep Learning (PIDL) framework. The framework integrates an optimized Joseph's instantaneous point source model, which provides a rapid baseline prediction incorporating radionuclide decay and atmospheric deposition, with a Denoising Diffusion Probabilistic Model (DDPM). The DDPM is uniquely employed to learn the complex distribution of the physical model's residual error, effectively reconstructing a high-fidelity concentration field from a biased initial prediction. To comprehensively validate the model's performance and generalization capabilities, we utilized two distinct and representative radionuclide release scenarios: the acute, high-concentration release from the Fukushima Daiichi Nuclear Power Plant (FDNPP) accident and the chronic, low-level discharges into the Irish Sea from the Sellafield facility. The results demonstrate that the PIDL model significantly outperforms the standalone physical model. On the Fukushima dataset, the coefficient of determination (R2) for 137Cs concentration prediction improved from 0.78 (physics-only) to 0.96 (hybrid model), with a corresponding 68% reduction in the Root Mean Square Error (RMSE). Critically, predictions are generated within seconds, meeting the stringent time constraints of emergency response. Furthermore, this study establishes a clear framework linking the model's rapid predictions to the principles of Probabilistic Safety Assessment (PSA) and specific nuclear emergency response actions. The model's outputs provide dynamic, data-driven inputs for consequence analysis, supporting dynamic hazard zone mapping, optimization of monitoring resources, and public communication strategies. This work presents a computationally efficient tool designed to support decision-making in the critical early stages of a coastal nuclear emergency.