AI-driven real-time control of radiological contamination and personnel dose in nuclear transient scenarios

This study presents an artificial intelligence (AI) system designed to optimize emergency responses during localized nuclear incidents involving containment breaches and gas filtration failures. The system integrates a feed-forward neural network (ANN) with optimization algorithms, including fmincon and ParticleSwarm Optimization (PSO), to manage critical parameters such as ventilation speed, activation timing, and personnel evacuation. A synthetic dataset of 3600 samples, simulating the dispersion of gamma and beta radiation, was generated using MATLAB. The ANN achieved a coefficient of determination (R²) of 0.9876 and a root mean square error (RMSE) of 0.3421. The AI system reduced radioactive concentration by 15 % and cumulative effective dose by approximately 30 % (1.8 mSv vs. 2.5 mSv) compared to a baseline scenario (fixed ventilation at 2 m/s, evacuation at 30 min), representing the minimum operational ventilation speed per Table 3, chosen to represent a conservative, non-optimized response typical of manual procedures (IAEA, 2018). Personnel exposure was maintained at 1.8 mSv over 30 min, well below the IAEA occupational limit of 20 mSv/year (annual context only), with decision cycles completed in under 3 s. This framework demonstrates the feasibility of AI-driven real-time nuclear emergency management, offering improved decision-making over traditional methods while ensuring compliance with IAEA GSR Part 3 and Part 7 standards.