Convolutional Neural Networks for MagLIF Stagnation Image Applications

Magnetized Liner Inertial Fusion (MagLIF) is a magneto-inertial fusion (MIF) concept being studied on the Z-machine at Sandia National Laboratories. MagLIF experiments employ imaging diagnostics that are used to constrain fuel plasma conditions and morphology. Analysis of these images involve tasks such as region of interest selection, background subtraction, and image registration of multiple images. While in principle these are not too difficult for an expert, manual treatment is tedious, increases risk of irreproducibility, and impedes quantification of uncertainty. To reduce the need for user input and time required to achieve common image analysis tasks, we present a convolutional neural network (CNN) based image segmentation able to detect pixels belonging to the stagnation column in Crystal X-ray Imager (CXI) images. This enables more fully automated image analysis pipelines and collective assessments of many images, which may lead to physical insights. In particular, we utilize this CNN approach to demonstrate an automated background subtraction pipeline. This enables statistical analysis of both slowly varying background signal and random noise in a large ensemble of CXI images, which will lead to accurate models of background and noise that are commensurate with experiment. In addition, we conduct unsupervised clustering of background subtracted images using an image similarity metric. We look for physically meaningful clustering to assess the viability of chosen metric for image-to-image comparison.