An autoencoder based reduced order model of low density plasma for optimal experimental design

Parasitic losses at Sandia's Z Machine may occur through near vacuum conditions and limit the current supplied to the load. A low density plasma experiment, consisting of a cylindrical target with an evacuated interior, has been developed to study these losses. However, due to the expense of shots at Z, experiments must be selected carefully to optimize information gain. Simulations aid experimental design, but due to high computational costs, reduced order models (ROM) are necessary. We introduce an autoencoder based ROM for this low density plasma experiment and apply it towards experimental design. Our ROM consists of two parts, an autoencoder that encodes the MHD fields in a latent space, and a ResNet that evolves the latent variables in time. This architecture is trained on 2D GORGON simulations of the low density plasma experiment for a variety of vacuum floors and geometries. Using this ROM, we select geometric parameters with different fixed vacuum floor settings that maximally distinguish quantities of interest, the dynamics of the magnetic fields and densities at various locations.