Abstract
Designing a tokamak fusion reactor requires exploring a high-dimensional parameter space where multiple coupled variables must simultaneously satisfy physics constraints and engineering limits. Conventional exhaustive search approaches become computationally prohibitive as dimensionality increases. This paper demonstrates the application of multi-objective Bayesian optimization (MOBO) to tokamak design space exploration using a low-fidelity system analysis code.We compare state-of-the-art acquisition functions against an exhaustive search baseline for bi-objective optimization minimizing major radius (a capital cost proxy) and maximizing net electric power. Results show that all MOBO methods recover Pareto fronts comparable to exhaustive search while requiring orders of magnitude fewer function evaluations. We investigate the effects of acquisition function choice, batch size for parallel evaluation, and constraint handling strategies on convergence behavior. The methodology provides practical guidelines for efficient design optimization and establishes a framework applicable to higher-fidelity,computationally expensive plasma simulations.

This publication has 25 references indexed in Scilit: