A Machine-Learning Approach to Stellarator Plasma Confinement Design and Optimization in Fusion Reactors

Abstract

This thesis aims to assess the viability of neural networks as a method of predicting steady-state plasma equilibrium solutions for magnetic field plasma confinement within a fusion reactor. The approach to this problem is broken down into several stages. The first stage involves the development and investigation of the performance of a Multilayer Perceptron (MLP) neural network tasked with producing axisymmetric equilibrium solution magnetic axis locations to be used in the computation process of the DESC equilibrium solver software. In the second stage, the task of the neural network is expanded to include the prediction of complete axisymmetric equilibrium solutions. The results of this stage motivated a change in approach for next stage, involving the development of a neural network that can predict stellarator equilibrium solutions. In the final stage, by considering the characteristic properties of stellarators, the task of the neural network is reduced to the prediction of partial solutions for stellarator equilibrium. The progression of the project stages is intended for a thorough analysis of neural network learning performance and capabilities, given the complex nature of mapping between the input boundary conditions and the equilibrium solutions. The identification of an optimal, successful neural network approach to equilibrium solution generation will significantly improve the efficiency of the plasma equilibrium solution generation process, which will increase the feasibility of future optimization for various objective functions. The results of this study suggest that the most promising approach to the utilization of neural networks in this context is to make accurate partial equilibrium solution predictions that can then be used to compute the full solution.