Abstract Details
Abstracts
Author: Jesse Viola
Requested Type: Poster
Submitted: 2026-03-19 13:35:21
Co-authors: T.M.Schuett, S.A.Henneberg
Contact Info:
Massachusetts Institute of Technology
175 Albany St
Cambridge, MA 02139
United States
Abstract Text:
We present an automated hyperparameter optimization framework for stellarator coil design using Bayesian optimization with Gaussian processes. Built around the stellarator optimization framework SIMSOPT [1], the framework efficiently searches an 11-dimensional parameter space governing multi-objective coil optimization, including penalty targets and weights for curvature, torsion, electromagnetic forces, and geometric constraints, along with Fourier order and iteration budgets. Given a performance metric f (w) built from a linear combination of each penalty and hyperparameter space W, the framework solves min_(w∈W) f(w) through two-phase optimization: random exploration followed by Bayesian-guided search using an Lower Confidence Bound (LCB) acquisition function.
Applied to a quasihelically (QH) symmetric stellarator with finite-beta equilibria, the framework achieves a 45% improvement in combined multi-objective score compared to random search (f(w) = 1098 using Bayesian optimization versus f(w) = 1979 using a random search) in the same number of iterations (N = 30). The score improvement is primarily due to a dramatic reduction in torsional strain (186.9 m−1 to 54.0 m−1) as well as improved minimum coil-surface clearance (71.8 cm to 73.6 cm), while maintaining comparable field error (2.19% versus 2.16% |B · n|/B).
[1] Landreman et al., (2021). SIMSOPT: A flexible framework for stellarator optimization. Journal of Open Source Software, 6(65), 3525.
Characterization: 6.0
Comments: