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

Area of Science:

• Nuclear Engineering
• Environmental Science
• Artificial Intelligence

Background:

• Assessing radionuclide dispersion in coastal waters during nuclear accidents is critical for emergency response.
• Conventional simulations face limitations in speed and predictive accuracy.
• There is a need for rapid, high-fidelity tools for real-time decision-making.

Purpose of the Study:

• To develop an innovative hybrid Physics-Informed Deep Learning (PIDL) framework for radionuclide dispersion assessment.
• To overcome the latency and fidelity limitations of existing marine dispersion models.
• To provide a computationally efficient tool for coastal nuclear emergency response.

Main Methods:

• Integrated an optimized Joseph's point source model with a Denoising Diffusion Probabilistic Model (DDPM).
• The DDPM learns the physical model's residual error for high-fidelity reconstruction.
• Validated using Fukushima Daiichi Nuclear Power Plant (FDNPP) and Sellafield discharge scenarios.

Main Results:

• The PIDL model significantly outperformed the standalone physics model.
• Achieved R² of 0.96 for 137Cs prediction (vs. 0.78) with a 68% RMSE reduction for Fukushima data.
• Generated predictions within seconds, meeting emergency response time constraints.

Conclusions:

• The hybrid PIDL framework offers rapid and accurate radionuclide dispersion predictions.
• The model supports Probabilistic Safety Assessment (PSA) and emergency response actions.
• Provides dynamic inputs for consequence analysis, hazard mapping, and resource optimization.