In neutral-particle transport shielding problems, variance-reduction methods are used in Monte Carlo (MC) simulations to bias the progression of tracked particles toward user-defined detectors or regions of interest. These biasing techniques allow for converged results in areas that would otherwise be poorly sampled due to low neutron or gamma fluxes relative to the fixed source. One widely used state-of-the-art methodology in shielding simulations is the Consistent Adjoint-Driven Importance Sampling (CADIS) method, which is a hybrid transport methodology that uses deterministic adjoint solutions to define weight window (WW) targets for particle splitting, rouletting, and source biasing during MC. However, for large problems, the WW data can require prohibitively large amounts of memory (tens to hundreds of gigabytes). This can make the simulation not feasible with the available computational resources, or it can restrict execution to a small fraction of nodes with large enough memory, thus significantly reducing the available resources and increasing the turnaround time needed to complete intended analyses.

A novel methodology and data structure have been developed and implemented within the MONACO and MAVRIC sequences of the Scale 6.1 code package that greatly reduces memory requirements for storing WW maps by orders of magnitude. The data structure is accompanied with an algorithm that determines mesh reduction through coarsening and refinement using contributon response theory. Large memory savings are achieved by using separate block-structured grids for each energy group. The implementation of this methodology leads to a fractional increase in biased MC simulation time due to tracking particles through a more complex data structure storing the WW targets. For large shielding problems, enhanced parallelism enabled by memory reduction more than compensates for the decline in biased MC performance resulting in an effective speedup in solution time. Here, the improvements and drawbacks in the methodology are demonstrated on the relatively small but well-known Pool Critical Assembly shielding benchmark. The methodology showed a reduction in memory of from 163 to 194 times, with only a limited slowdown in biasing efficiency between 1% and 9%.