(A) study of distinguishing wall heat transfer mode of MARS-KS using unsupervised learning

Abstract

A nuclear system analysis code has been developed to perform realistic multi-dimensional thermal hydraulic system analysis over the years. The MARS-KS is an established multi-dimensional thermal hydraulic code used for the analysis of reactor transients, simulation of experiment, various licensing activities, as well as safety research. The system analysis code encompasses diverse empirical correlations, not only limited to governing equations and constitutive equations. Wall heat transfer is a type of constitutive equation that requires 10 different thermal hydraulic variables, and distinguishing wall heat transfer mode depends on the code developer logic using specific variables, and critical heat flux (CHF) tables. Despite improvements in this logic through continuous research are ongoing process, there still remains challenges in effectively dealing with multi-dimensional variable problems. Unsupervised learning has its strengths in classifying high-dimensional variables. In addition, there is an advantage that the performance can be augmented with accumulation of experimental data. In this study, nucleate boiling, transition boiling, film boiling data were generated and their wall heat transfer modes were classified with unsupervised learning. Validation of trained unsupervised learning was performed using real experimental data. The CHF region was identified using unsupervised learning and was compared with the values obtained using the CHF lookup table to evaluate the potential of unsupervised learning method applied in this area.