Heuristic optimization of group structure using Physics-Based Fitness Approximation

In Multi-Group (MG) deterministic radiation transport simulation applications, Fine-Group (FG) cross section libraries are often collapsed, or down-sampled into Broad-Group (BG) libraries to reduce computational cost. Previously, we developed a heuristic approach to determine the optimal BG group structure using the Particle Swarm Optimization (PSO) algorithm. A key step in PSO is to evaluate the fitness function, which measures the optimality, or fitness of a given group structure. In this paper, we demonstrated that a Simulation-Driven Fitness Function (SDFF) can be readily formulated by performing transport simulations on a BG structure, and then comparing the results, such as flux distribution, detector response, and/or k-effective, with the reference FG simulation solution. While the PSO approach is very effective in searching for global optimum in a very large solution space, the SDFF evaluation process can be very time-consuming since it requires a large number of BG transport simulations, even though a single BG simulation usually costs much less time than the reference FG calculation. To overcome this problem, a novel Physics-Based Fitness Approximation (PBFA) approach is developed based on the contributon transport theory. Compared to SDFF, PBFA does not require BG transport simulations. Therefore, it significantly reduces the computational cost for evaluation of the fitness function. We performed a thorough numerical test on a neutron detector model, the Dual Range Coincidence Counter, to examine the efficacy of PBFA. We found that a strong positive correlation exists between SDFF and PBFA. Thanks to this new finding, we can replace PBFA with SDFF to accelerate the PSO algorithm for the group structure optimization application.