Variance Reduction

General considerations

The time dependent tracking of the neutron population in a multiplying, near critical medium is very challenging in terms of Monte Carlo convergence. A naive analog game in most cases would statistically diverge, moreover it will give an underestimate of the power as the very low chance contributions of a high number of fission in certain chains see Fig. 7. Therefore the calculation is performed always keeping a single particle as a sample of the neutron population gaining or loosing weight at interactions. The neutron weight distribution must be kept around the mean for ensuring statistical convergence. The neutrons are followed from time interval to time interval and the population at the interval ends using splitting and Russian roulette while keeping the total population number constant. Having single, non-branching calculations also supports the architecture of the GPU where threads can be set to single neutron chains.

Fig. 7. Analog and non-analog simulation results for time dependent power evolution for a multiplying medium. Analog simulation produces an underestimate of the power

Biased sampling schemes are applied at fission yield, delayed neutron, interaction type sampling with ongoing development regarding path length sampling and angular biasing.

Optimization of the GPU workflow

The key idea of GUARDYAN is a massively parallel execution structure making use of advanced programming possibilities available on CUDA enabled GPUs. In order to maximize performance however, the architecture calls for some major deviations from a traditional MC code.