Source term inversion of nuclear accident with random release durations based on machine learning

When a nuclear accident occurs, a large number of radioactive nuclides are released into the environment, seriously affecting the environment and human health. Machine learning-based source term inversion addresses the limitations of traditional methods, such as the dependence on a priori information and the limited adaptability to different scenarios, which are crucial for assessing the consequences of nuclear accidents and supporting emergency decision-making. In this study, a source term inversion model, which was developed based on a long short-term memory (LSTM) neural network and can reduce dependence on prior information, was proposed to determine the release rate under conditions of an unknown release duration during a nuclear accident. The model was optimised using the Optuna algorithm and tested under a nuclear accident scenario in which three nuclides, Kr-88, Te-132, and I-131, and the meteorological conditions were changed once. The results showed that the mean absolute percentage error (MAPE) of the release duration prediction obtained using the proposed model was less than 20 % when the release duration ranged from 1 to 120 min. The MAPEs of the Kr-88, Te-132, and I-131 release rate predictions were 19.68 %, 10.83 %, and 28.63 %, respectively, when the release duration ranged from 1 to 30 min. Stability analysis experiments show that while the gamma dose rate remains the most critical input to the source term inversion, the release duration has the highest sensitivity to it. Except for wind and rainfall noise, the noise of the remaining meteorological conditions has very little effect on stability, which emphasizes the robustness of the model under typical meteorological variations.