Machine Learing-Boosted Diagnostics Using Autoencoders

While plasma simulations provide a complete picture of the plasma state, the same cannot be said about measurements. To better perfect our diagnostic tools, synthetic diagnostic data can be calculated using the expected response of various sensors using the simulation data. But in order to find a simulation that reproduces the measurements, computationally expensive trials may be required.

This work aims to explore how autoencoders, a type of generative machine learning model, can be used to map synthetic diagnostics signals to underlying simulation data. Autoencoders map their inputs to a low-dimensional latent space. They then reconstruct the original data from the latent space representation. During training, the latent space adapts to represent the most important properties of the training data set. Our autoencoder is trained on reconstructing pairs of simulation and synthetic diagnostic data. The simulation data was produced using the Gkeyll code and a simplified gas-puff imaging diagnostic is used as the forward model.

We explore how the dimensions of the latent space relate to various features of the training data set. Additionally, we explore to what degree it is possible to reconstruct both simulation data and synthetic diagnostic data when parts of the input are missing.