Harnessing Multi-fidelity Design, Analysis, and Optimization Techniques to Advance Fusion Reactor Design

The design of fusion reactors necessitates a comprehensive assessment of interacting components to ensure engineering and economic feasibility. Complex systems, like fusion power plants, are often preliminarily modeled using low-fidelity ‘systems codes’ for efficient design space exploration. This project investigates applying aerospace-derived multi-fidelity techniques to incorporate higher-fidelity models while expediting system convergence and optimizing plant design. Multi-fidelity analysis utilizes high-fidelity models to locally calibrate low-fidelity models. We automate the integration of high-fidelity and low-fidelity models using Gaussian process regression, a Bayesian machine learning method. Low-fidelity power law models are ‘trained’ using high-fidelity models. Gaussian Process Regression is used to estimate the error of the low-fidelity model as design points vary from the initial training data set. If the error is large for a given design point, the point is added to the training set for the high-fidelity model, and the system is re-evaluated. We present progress on this work.