High-Fidelity PBR Depletion Calculation in Griffin using Artificial Neural Network
Author: Arif Akhter Rangon, Texas A&M University
Abstract: Serpent-2 Monte Carlo code has recently been so popular for transport calculations of pebble bed reactors (PBRs) due to its capability to account for double heterogeneity and stochastic explicit geometry for pebbles. In addition, a novel deterministic method for PBR calculation is Pebble Tracking Transport (PTT) algorithm which is implemented in Griffin, a fl agship code developed by Idaho National Laboratory. Griffin-PTT can do full core PBR transport calculation at pebble-level if it is provided with macroscopic homogenized cross sections for each pebble. In this work, torch-scripted Neural Networks have been interfaced with Griffin-PTT that predicts the macroscopic cross section for each pebble on the fl y based on the individual pebble spectrum provided by PTT. The NN model is trained from the infi nite lattice cell calculation results from Serpent for single pebbles with an additional buffer layer surrounded by the pebble which contains the homogenized nuclide density of a pebble. By increasing the buffer thickness, we generated the training dataset (neutron spectrum and cross section sets) for the NN. Using this NN in Griffin, we compared the results of a shortstack PBR core with 22930 pebbles that yields only 217 pcm eigenvalue difference compared to Serpent reference calculation and the individual pebble power error stays below 8%. Future work includes the depletion calculations with temperature and burnup dependencies to create a more robust NN that can predict cross sections on the fl y for each isotope of interest at pebble-level, such as Plutonium and higher actinides, based on the neutron spectrum, pebble burnup and temperature.