High-precision global tallying for Monte Carlo neutron transport simulation using smaller samples by denoising with cascaded neural network

The global tallying problem of Monte Carlo particle transport simulation has wide applications in the nuclear engineering field, inertial confinement fusion field, and so on. Here, tally means any physical quantity we concerned and global tallies mean all tallies in grids of the whole geometry space or phase space. Generally, global tallying problem is more difficult than seeking single target tally and often needs numerous samples to get precise-enough global solutions, thus increases the time cost greatly. This paper proposes a novel idea to get precise global tallies by relatively fewer samples, thus increases the global efficiency to some high level. This strategy relies on a cascaded neural network to remove the global and local noises accompanied with all tallies. The Kobayashi model is utilized to show the correctness and efficiency of this method. The simulation of two-dimensional C5G7 model, which belongs to deep penetration problem, reveals this method can indeed improve global precision by up to $128$ times with additional time cost less than one second. As a popular viewpoint, the deep penetration problem is considered as one of the most difficult challenges in Monte Carlo particle transport simulation. It can be said safely this method has non negligible value in many situations, such as the nuclear reactor global physical field calculation or the calculation of global feedback quantities caused by particle transport in multi-physics coupling calculation.