Predicting and Controlling Plasma Profiles of Tokamaks through an Autoencoder System

Abstract

The DIII-D National Fusion Facility is part of the ongoing effort to achieve magnetically confined fusion and establish the scientific basis for the optimization of the tokamak approach to fusion energy production. Because the experiments performed require time and material resources, development of computational models to recreate experiments in the DIII-D tokamak are invaluable and a source of ongoing research. One such model to simulate the evolution of plasma profiles over time is explored in this thesis. In this project, an existent linearly recurrent autoencoder neural network is modified to test its performance on predicting plasma profiles 200 milliseconds into the future, given the relevant actuator inputs and initial plasma state. Data collected from the 2010-2018 experimental campaigns on DIII-D tokamak is used to train and test the model. Model predictive control is implemented on the linearized model learned by the system to find an optimal set of actuator inputs which drive the plasma from an initial state to a predefined target state. Finally, the performance of the controller is then evaluated by propagating its output through both the autoencoder model and a convolutional neural network designed to make accurate predictions on the same timescale. Both the controller and the autoencoder model are found to make reliable forecasts of actuator inputs and plasma profiles respectively, and therefore may be utilized as part of a larger system to simulate plasma physics experiments in the future.