Machine Learning Approach to Analyzing Data in Japanese BWR Chemistry Database

Abstract Understanding the relationship between the dose rate of the primary loop recirculation (PLR) piping and chemistry parameters is the beginning of developing the techniques of reducing the radiation exposure in boiling water reactors. In this study, a regression model was built using a machine learning method, gradient boosting, to identify the relationship between the dose rate of the PLR piping and several chemistry parameters. Categorical Boosting, EXtreme Gradient Boosting, and Light Gradient Boosting Machine were used to build a regression model, with the dose rate of the PLR piping as the target variable and chemistry parameters as the explanatory variables. The coefficient of determination and root mean squared error across these three models are nearly identical, demonstrating the similar accuracy of different gradient boosting methods. SHapley Additive exPlanations values were used to interpret the internal structure of these models and rank the importance of parameters associated with the dose rate of the PLR piping. All three models highlighted the same top-tier parameters, indicating that comparing regression models could be beneficial for enhancing the robustness of parameter extraction. Only within certain ranges was a strong relationship found between the dose rate of the PLR piping and these chemistry parameters, implying non-linear relationships and optimal ranges corresponding to lower values of the dose rate of the PLR piping.