Hybrid Deep Learning-Metaheuristic Integration for Contaminated Area Prediction and Source Term Inversion of Nuclear Energy Safety

摘要: Nuclear energy safety is indispensable to balancing the dual goals of harnessing nuclear energy and mitigating the risks of radioactive leakage, which is critical to global energy transition and environmental protection efforts. A Hybrid deep learning approach combined with metaheuristic algorithms for predicting contaminated areas and inverting source term parameters in nuclear explosion models was designed, and a metaheuristic algorithm was used to invert the source term parameters of nuclear explosion. The physical processes of particle distribution, deposition mechanisms, radioactive activity, and biological dose resulting from nuclear explosion radioactive clouds were simulated. The relationship between explosion parameters and contaminated areas is systematically analyzed. Subsequently, radiation safety standards are employed to quantitatively evaluate contamination levels. A hybrid deep learning methodology is then developed to construct a rapid prediction model for contaminated areas. Finally, metaheuristic algorithms are utilized to invert explosion source term parameters, with comparative analysis of different algorithms’ stability and accuracy. Results demonstrate that the integration of Bayesian optimization (BO) and attention mechanisms significantly enhances model prediction accuracy. Metaheuristic algorithms exhibit fast convergence speeds and excellent search capabilities in inverting nuclear explosion source parameters. This research contributes to enhancing the effectiveness of nuclear emergency response actions, provides a scientific basis for disaster assessment and rescue operations in nuclear environments, and offers a robust technical toolkit for safeguarding nuclear energy safety in extreme accidental or intentional nuclear event scenarios.