Project Description

Neutrinos are fundamental particles that may hold the key to understanding why there is a preference for matter over antimatter in the makeup of the universe today. The 2015 Nobel Prize in Physics was awarded for the discovery of neutrino oscillations, which shows that neutrinos have mass. We seek to establish whether neutrinos and antineutrinos oscillate differently and carry out a precision neutrino-nucleus cross section program.
The 2014 Particle Physics Project Prioritization Panel report highlights the physics of neutrino mass as one of the five Science Drivers for High Energy Physics (HEP). This proposal addresses this Driver with a synergistic neutrino cross section and oscillation physics program. A recent revolution in machine learning powered by new Graphics Processing Units and deep learning algorithms has propelled computers past humans in certain pattern recognition exercises, particularly in computer vision. Modern detectors are effectively imaging devices, and early results indicate that deep learning will significantly improve neutrino experiments by improving reconstruction efficiency and widening the set of accessible event topologies. Furthermore, the massive simulations available in HEP make this an excellent arena to study deep neural network optimization and subjects such as representation transfer and semantic segmentation in a physics context.
This proposal will analyze two complementary neutrino detectors, MINERvA and NOvA, and use deep learning to drive improvements in their physics reach. We will optimize deep neural network performance by conducting an enormous parameter search, using an evolutionary algorithm to evolve solutions using custom software: the Multi-node Evolutionary Neural Networks for Deep Learning (MENNDL) package. We will leverage common simulation software and neutrino interaction processes while contrasting different detectors to analyze network optimization. Addressing all of these questions is only possible with leadership computing.

Allocation History

Source Hours Start Date End Date