A Convolutional Neural Net (CNN) was trained to determine the optimal decomposition ofa voxel domain for Monte Carlo electron-photon radiation transport. The training database
was developed by collecting photon flux and surface current tally data for a simple shielding
problem and a simplified human phantom brain model. The voxel matrix inputs were mapped
to the tally outputs by the CNN, allowing the CNN to accurately predict tally results from
material and source data. The predicted flux was then used to determine the shape and size
of the subdomains, and to calculate the size of the ghost zones. The intent for the CNN is to
reduce the amount of trial and error that is often necessary to decompose domains manually
by a user. Poor domain decomposition can lead to longer run times. Furthermore, this work
is meant to serve as a proof of concept for more complex geometry types and applications.
The CNN predicted decomposition performed at best 1.6x than other decomposition schemes
for the shielding problem and 1.3x for the human phantom problem. This process was done
to demonstrate the power of intelligently choosing subdomain boundaries rather than using
arbitrarily chosen boundaries.