Developments in Domain Decomposition for Neutron Transport Problems on High Performance Computing Systems

Domain decomposition is a technique used to reduce memory overhead on large neutron transport problems. During the Center for Exascale Monte Carlo Neutron Transport project, this technique was developed in two major thrusts: processor allocation, and reproducibility. Processor allocation on high performance computing systems was improved through the development of a machine learning model, which can predict the computation load of a decomposed problem’s subdomains quickly and accurately. Reproducibility was improved via a new domain decomposition methodology, which takes advantage of a particle-tracked random number generation method to ensure domain-decomposed simulations generate identical results to non-decomposed simulations. Applications of both of these methodologies in the Monte Carlo Dynamic Code solver are discussed, with a focus on the 3D C5G7 eigenvalue problem.