Efficient Representation of Proton Images for Machine Learning Applications in Laser Plasma Interactions

In HED experiments, strong magnetic fields are generated in laser-plasma interactions. Probe protons are deflected in the magnetic field as a result of the Lorentz force and these protons arrive at an image plate and are captured. These proton radiograph images produce observable features that aid in optimizing ICF experiments and improving implosions. However, these images are unable to provide the proton trajectories and the shape of the magnetic field. Therefore, a method that would be able circumvent this would be using machine learning applications to train a neural network to guess the correct magnetic field, given the proton image.

In order to make machine learning much more feasible, our large data set of images requires efficient representation to facilitate machine learning analysis. Our research employs advanced image compression techniques and the singular value decomposition (SVD) to extract key spatial patterns and reduce data dimensionality to improve data representation for ML algorithms. We present numerical analyses of the transformed data to evaluate information storage requirements needed. This work aims to advance understanding and reconstruction of underlying path-integrated magnetic deflections in HED experiments, while exploring the application of neural network-based methods.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract No. DE-AC52-07NA27344