Machine Learning for Magnet and Coil Design
Surrogates and optimizers explore coil geometries and current sets that produce a target magnetic field while respecting stress and superconductor limits.
The coil design problem
Magnetic confinement depends on precisely shaped fields produced by coils. Designing them means choosing coil positions, shapes, and currents so the field meets confinement requirements while stresses, forces, and superconductor operating limits stay within bounds. The design space is large and the constraints are coupled.
Surrogates for the physics
Evaluating a coil set means computing the magnetic field, the resulting plasma equilibrium, and the mechanical stress. Each is expensive. Learned surrogates approximate these maps so an optimizer can evaluate thousands of candidate coil sets quickly, reserving full electromagnetic and structural solvers for finalists.
- Field-shape surrogates mapping coil currents to the resulting flux geometry
- Stress surrogates estimating peak conductor and structure loads
- Superconductor margin models relating field and temperature to critical current
Optimization
With fast surrogates in the loop, multi-objective optimization balances field accuracy against stress and stability margins. Because coil design trades many competing goals, the output is a set of Pareto-efficient candidates rather than a single answer, which engineers then examine in detail.
High-field context
Modern concepts push high fields with rare-earth-barium-copper-oxide superconductors, where operating margin is tight. The Hyperion breeder concept targets a peak field of 16.84 tesla at the coil with roughly 8 tesla on axis, a regime where accurate stress and margin surrogates are essential to keep candidate designs feasible.
All such figures are simulation targets for a concept under design. Surrogate-driven exploration accelerates the search, but final coil designs are validated with full electromagnetic and structural analysis, and no magnet performance is claimed for hardware not yet built.