Rapid and accurate inversion of radioactive source terms following a nuclear accident is critical for risk assessment and emergency response. Static neural networks are widely used for inversion because of their fast inference, but they struggle to capture the temporal dynamics of evolving source terms. To overcome this, we propose a source term inversion model that integrates a Gated Recurrent Unit (GRU) with Optuna’s automated hyperparameter optimization framework (Optuna–GRU). The model leverages simulated environmental radiation monitoring data from three consecutive time steps to predict source strengths at the next two steps. Trained and validated on data generated by the enhanced Gaussian Plume Model (GPM+), the model achieved a MAPE as low as 11.95% and an R2 of 0.9705 at the first prediction step, with R2 remaining high (0.9344) at the second step. Training required only 510.70 seconds, and inference was completed in under 1 ms, enabling real-time application. Sensitivity analysis revealed that performance is most affected by wind speed perturbations and far-field monitoring concentrations, while showing robustness against precipitation and release height uncertainties. By integrating Optuna, the method improves both network structure optimization and training efficiency, achieving high accuracy and strong extensibility.
Zhao et al. (Sun,) studied this question.