Application of artificial neural network for assembly homogenized equivalence parameter generation

Abstract

The development of a new reactor core and optimizations for fuel assembly (FA) structures necessitates the generation of homogenized equivalence parameters, including macroscopic cross-sections (XS), for numerous FA configurations. A feed-forward convolutional neural network, XSNET, has been created to streamline this process. Currently, XSNET can produce these parameters for fresh and burned UO2 fuel, both with and without Gadolinia (Gd) burnable absorber rods. In this study, we integrated XSNET with our in-house nodal diffusion code, RAST-K, to create a conventional two-step approximation code system. This system was tested on various tasks, including analyzing a 3-dimensional pressurized water reactor employing 16 x 16 FA types with Gd rods. Over 1,000 unique loading patterns were considered for nodal calculation, which involved steady-state depletion calculations and pin power reconstruction. The results obtained using XSNET/RAST-K demonstrated good agreement with the reference STREAM/RAST-K code system.