Risk Prediction and Mitigation of Drone-Deployed Radiological Dispersal Devices Using Physics and Machine Learning

The deployment of radiological dispersal devices (RDDs) via drones presents a novel security challenge, necessitating advanced tools for consequence assessment and response planning. We developed an integrated framework combining physics-based dispersion modeling, constrained optimization, and machine learning to evaluate such threats. Using a Monte Carlo approach, 2000 synthetic scenarios were generated incorporating five radionuclides (Cs-137, I-131, Co-60, Sr-90, and Am-241), meteorological variability, and geospatial risk zones. A constrained optimization routine based on the Limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithm with bound constraints (L-BFGS-B) identified adversarial scenarios that maximize contaminated area (>10 km²) while minimizing energy use and detection risk, revealing nonlinear trade-offs between dispersal effectiveness and operational stealth. Consequence modeling with Health Physics Code (HotSpot) and Java-based Real-time Online Decision Support system (JRODOS) showed systematic differences, with HotSpot predicting higher total effective dose (TED) and time-integrated air concentration (TIAC). I-131 posed the greatest acute thyroid risk, whereas Am-241 dominated long-term exposure. Protective action analysis demonstrated that reinforced sheltering reduces cumulative dose by up to two orders of magnitude compared to outdoor exposure. Finally, the machine learning framework achieved accurate and rapid predictions (R² = 0.975), with distance as the dominant predictor. These findings provide actionable guidance for emergency preparedness against drone-based RDD threats.