American Nuclear Society
Home

Home / Publications / Journals / Nuclear Science and Engineering / Volume 197 / Number 8

Using a Random Forest Model to Choose Optimized Group Structures

Thomas G. Saller, Vishnu Nair, Andrew Till, Nathan Gibson

Nuclear Science and Engineering / Volume 197 / Number 8 / August 2023 / Pages 2117-2135

Technical papers from: PHYSOR 2022 / dx.doi.org/10.1080/00295639.2022.2133940

Received:July 6, 2022
Accepted:October 6, 2022
Published:July 7, 2023

It is challenging to select an appropriate group structure for any given multigroup neutron transport problem. Many group structures were designed long ago, and the reasoning behind the creator’s choices may be unknown. In this work, we apply the simulated annealing optimization method to develop improved group structures for a set of test problems. We then use a random forest (a machine learning method) to identify which group structure will be the best for any new problem based on input characteristics, such as geometry and isotopics.

Simulated annealing spans a large solution space before narrowing in on an optimal solution, avoiding local minima by jumping around. Our solution space, however, is large and inconsistent, making finding the optimal group structure infeasible. Instead, we find potentially optimal group structures, ones that yield more accurate solutions than our standard group structures, but are probably not the “best” possible. Group structures are obtained for six classes of problems, ranging from a fast 233U system to a thermal 239Pu system. These were chosen to encompass a series of critical assemblies from the International Criticality Safety Benchmark Evaluation Project (ICSBEP) handbook. These optimized group structures were used in PARTISN for a large range of ICSBEP critical assemblies and compared to the traditional Los Alamos National Laboratory group structures. Our reference solution was from 618-group PARTISN runs. The results were used to train a random forest regressor model with bagging, which was then tested on similar benchmarks. The bagging regressor model chose the best group structure from 52% to 65% of the time, and a subjectively “good” group structure up to 91% of the time.