A two-stage gradient boosting framework coupled with Lagrangian particle tracking for rapid atmospheric radionuclide dispersion prediction in nuclear emergency response.

Predicting the atmospheric dispersion of radionuclides is central to nuclear emergency response, yet any useful prediction tool must balance physical fidelity against the speed that real-time decision-making demands. This trade-off is particularly challenging under non-stationary meteorological conditions, where shifting winds produce complex plume geometries that steady-state models cannot adequately represent. In this paper, we develop a lightweight two-stage gradient boosting (TS-GB) surrogate coupled with Lagrangian particle tracking for rapid concentration-field prediction over dispersion scales of 50 km on flat terrain. Unlike deep-learning surrogates, the framework requires only modest training data and no specialized hardware. Training data are generated by a Lagrangian Stochastic Particle Model (LSPM), which shares its governing equations with FLEXPART, NAME, and HYSPLIT, across 78 meteorological scenarios derived from the Shenzhen National Basic Meteorological Station. Because concentration fields are inherently sparse (zero-to-non-zero ratios up to 19:1), the framework separates zero-region classification from concentration regression, achieving R2 = 0.996 and a zero-class F1 of 99.2% under steady-state conditions. A cohort-based temporal decomposition then extends the framework to non-stationary wind fields without retraining. Under cumulative wind-direction changes up to 726°, the surrogate maintains R2 > 0.93, whereas the Gaussian Plume baseline yields R2 < 0 under the same conditions. Residual errors concentrate along the low-concentration plume periphery and do not propagate into emergency decisions: when predicted concentrations are converted to ambient γ dose rates and classified against Operational Intervention Level (OIL) thresholds, decision consistency exceeds 98.9%, with zero missed detections under moderate and high release intensities. Prediction time is reduced from approximately 29 min to under 0.4 s, making the surrogate suitable for real-time emergency risk zoning.