Physics-informed neural networks for sparse neutron source reconstruction

Methods

Summary

This project will utilize the PyTorch framework to build a neural network that is trained to understand the physics of the detection system and its sensitivity in ∼ 2π. The neural network will be trained using random data generated with a proprietary methodology that preserves the physics of the system so that there is no biasing the model and tested with Monte Carlo generated data with GEANT4. The training will be accelerated on a Graphics Processing Unit (GPU) for speed and efficiency while taking into account memory utilization.

Challenges

The main challenge will be the virtual random access memory (VRAM) limit on the GPU and therefore compromises will need to be made between speed (data transfers from CPU to GPU) and bandwidth (how much training data can be used).

Pre Analysis Plan

The plan is to compare the trained model's one-shot prediction against a Maximum Likelihood Expectation Maximization (MLEM) model after convergence for speed, resolution (FWHM in azimuth and elevation), and quality (pixelization).