Abstract
Decision support in radiological emergencies requires dispersion forecasts, tail-risk dose estimates, and evacuation plans that balance conflicting goals. We present a framework that couples a convolutional and recurrent neural network with temporal attention (CNN–LSTM), used as a fast surrogate of a Gaussian plume model; a Conditional Value-at-Risk (CVaR) dose objective; five-objective evacuation optimization with the NSGA-III evolutionary algorithm; and four radiological indices. Each component is tested against an explicit reference. Trained on 40 years of ERA5 reanalysis (1985–2024), the surrogate achieved R² = 0.855 on a temporally independent test set (2019–2024) and, in a pre-registered test, reduced the mean absolute error relative to persistence by 17.2% (95% block-bootstrap confidence interval 15.6–18.7%). Against the Project Prairie Grass tracer measurements, the reference plume reproduced the mean concentration (geometric mean bias 1.04) but with large case-to-case scatter (fraction within a factor of two 0.475), and did not meet the pre-registered acceptance criterion. On a synthetic ten-zone planning area with objectives for evacuation time, dose CVaR, cost, logistic access, and collective panic, a pre-registered comparison over 30 independent runs ranked NSGA-III ahead of NSGA-II and a decomposition-based algorithm on two standard quality indicators (Wilcoxon p < 0.01), with a small Hypervolume margin over NSGA-II (0.8%). Under a single declared specification of 10,000 synthetic scenarios, the ranking of scenarios by the cognitive-bias coefficient was robust to its assumed weights (Spearman ρ ≥ 0.90). Motivated by the 1987 Goiânia accident, the framework is released with its code, pre-registrations, and an automatic check of every reported number against the analysis outputs.

This publication has 16 references indexed in Scilit: