A multi-model comparative study on container parameters in the activity inversion of Kr-85 in a closed container

When utilizing external dose rates to infer the Krypton-85 (Kr-85) activity within a sealed container, multiple factors can affect the accuracy of the activity inversion results. This study investigates the effects of multidimensional container parameters on the dose rate per unit Kr-85 activity and multi-model comparison to account for variations in multidimensional container parameters. A three-dimensional model was developed based on a practical sealed container using Monte Carlo simulation software, and a dataset of dose-rate response coefficient was created for multiple variables, including the container inner diameter, wall thickness, material, and internal pressure. The performance of five distinct machine learning algorithms in the inversion process was assessed: Ridge Regression (Ridge), Random Forest (RF), XGBoost, Multilayer Perceptron (MLP), and Support Vector Regression (SVR). The findings indicate that the SVR model demonstrates superior accuracy and stability in predictions at three different detection points outside the container in this study, with a coefficient of determination R² consistently exceeding 0.99. Sobol sensitivity analysis reveals that the container's inner diameter is the primary factor influencing the inversion results. The results demonstrate that the SVR model can effectively address the nonlinear inversion of radiation fields, highlighting the computational feasibility of the proposed approach for rapid and non-destructive activity inversion of Kr-85 in sealed containers.