Development of methodology for improving nuclear system analysis code using machine learning technique and non-parametric statistics theory

Abstract

To improve the accuracy of nuclear safety analysis codes, many studies on the improvement of the constitutive equations were conducted. However, there are only few studies on developing a methodology to improve the code accuracy by directly using integral effect test (IET) data. In this study, a methodology that can improve the accuracy of the code by directly using the IET data which is not easy to analyze is newly suggested and developed. First of all, the constitutive equation data that can cover a wide range of TH conditions are generated. Subsequently, the heat transfer/flow regime maps of the safety analysis code are subdivided. Instead of subdividing the regime map by setting arbitrary standards by the user, the characteristics of the data are found through data-driven modeling using artificial intelligence. The map is subdivided with the clustering indices of unsupervised learning algorithm. The multiplier coefficient is then applied to the subdivided regime to study the sensitivity and effect of the constitutive equations on the code prediction with respect to the TH conditions. The necessary number of multiplier coefficient sets are calculated from the Wilks’ formula which is based on the non-parametric statistics, and the multiplier coefficient is sampled with the Latin hypercube sampling method. The dominant sub-regime analysis, effect of single multiplier coefficient variation, optimal modification of constitutive equation are conducted for the IET experiment. The new methodology suggested in the thesis can be used to improve the code prediction results of complex IET data by identifying the direction for constitutive equations modification.