Nuclear accident source term inversion based on explainable machine learning methods

Abstract

Traditional source term inversion models rely on accurate a priori information as well as atmospheric dispersion simulations, leading to time-consuming source term inversion procedures. Previous studies have used machine learning (ML) methods such as neural networks to construct source term inversion models, which exhibit excellent inversion performance but usually lack model interpretability and have complex model structure and parameter tuning. To address this problem, an interpretable nuclear accident source term inversion model using ensemble learning combined with the SHapely Additive exPlanation (SHAP) method was developed in this study to estimate the nuclide release rate and the 2D location of the release point. In the model construction, Gaussian plume model is utilized to obtain data samples. To evaluate the adaptability of the model to accident scenarios, the model was trained under two types of accidents, known and unknown at the release point. The validity and accuracy of the model were assessed using statistical metrics, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute percentage error (MAPE), and mean distance error (MDE). The CatBoost model showed the best performance in both scenarios compared to the other three models. Model feature importance calculations and SHAP analyses revealed that the radioactivity concentration monitoring data had the greatest impact on the model inversion performance in both scenarios, and wind speed was an important parameter for this inversion model. Variations in meteorological parameters critically impair the reliability of source term inversion under unknown release scenarios.