Use of Generative Adversarial Neural Networks in Scattering Muography

Abstract

Many muography applications make extensive use of simulations to determine detector design or to train
imaging or regression algorithms. The computing cost of producing these simulations is usually quite high,
especially concerning the interaction of cosmic muons with matter. This work explores the possibility of
using Generative Adversarial Neural (GAN) networks to produce a fast and realistic simulation of the
multiple scattering process. The results of the network are confronted with GEANT4 simulations using a
benchmark problem related to the measurement of the inner wear of industrial pipes. The GAN is able to
reproduce the angular distributions and correlations with a speed-up factor of roughly 50 with respect to
GEANT4.