Novel Monte Carlo methods for proton transport simulations in the heterogeneous CPU-GPU environment

Abstract

A single event proton transport method has been developed using a cross section based, analog transport model. Cross section models were selected to achieve a balance of performance and accuracy for this study. Each method has been implemented in C++ and compared on the Snow HPC system at Los Alamos National Laboratory as well as on the Radiation Measurement and Dosimetry Group’s (RRMDG) graphics processing unit (GPU) cluster at Rensselaer Polytechnic Institute. The particular models being examined were verified and validated while restricted to the 1 MeV to 1 GeV energy regime, sufficient for proton radiography applications. Using an Nvidia K40c GPU, the single-event proton transport code was able to achieve 105 times speedup relative to serial MCNP6.2 simulations, equivalent to a 64 noderun on the Snow HPC machine with the MCNP6.2 code. The code developed was also able to successfully capture particle behavior from the Blur Test Object Experiment performed at Los Alamos National Laboratory proton radiography facility including the characteristic blurring and the radiographic phenomenon referred to as limbing. The single-event code is proven to be a powerful simulation tool well adapted to modern hardware, and capable of providing fast simulation results over small spatial scales to both desktop and HPC users.

High energy charged particles provide unique radiographic advantages when compared to neutral particles. Over the past two decades, the technology to direct and shape charged particle beams for radiographic imaging has become increasingly refined, permitting higher energies, up to 800 MeV at the Los Alamos Neutron Science Center, and finer spatial resolution in images. However a complementary increase in the quality of simulating these more powerful devices in radiation transport codes has been lacking. General purpose Monte Carlo transport codes rely on computationally efficient models for proton transport, with most models developed to suit the hardware of the early 2000s, in many cases sacrificing accuracy to achieve this. We propose a single event proton scattering algorithm of improved performance and comparable accuracy to contemporary Monte Carlo codes such as MCNP6 and Geant4. To understand and improve charged particle transport on next-generation computer architectures a traditional transport method has been translated from the MCNP6 code into C++ and the limitations of its underlying physics models have been documented. The physics evaluated were continuous energy loss, angular scattering, and energy straggling of protons.