Machine Learning for Coil-Current Optimization
Learned models find the coil-current set that produces a desired plasma shape and position, both offline for design and in real time for control.
Currents to shape
Given a coil configuration, the currents in each coil determine the plasma boundary shape and position. The inverse problem, finding currents that produce a target shape, must be solved offline during scenario design and continuously during operation as the plasma evolves.
Learned inverse maps
A network can learn the map from a desired shape descriptor to the required coil currents, trained on equilibria computed across many shapes. This gives an instant first solution that a feedback controller refines, and it provides a smooth starting point that avoids the local minima that trap iterative solvers.
- Inputs: target boundary points, elongation, triangularity, position
- Outputs: current in each poloidal field coil
- Constraints: current and voltage limits, force limits between coils
Real-time control
During operation the plasma shape drifts and must be held by feedback. Combining a learned feedforward current estimate with a feedback loop yields fast, accurate shape control. Reinforcement learning has been used to train controllers that directly command coil voltages to track a target shape.
Handling constraints
Practical coil control must respect current and force limits so coils are not overstressed. Constraints can be enforced by projecting network outputs onto the feasible set or by penalizing violations during training. Ignoring them risks commands the power supplies cannot deliver or that damage the structure.
Shape control is central to holding a designed configuration, such as the -0.30 triangularity target of the Hyperion breeder concept. In Kronos design work these controllers are developed and tested in simulation ahead of construction; they act on modeled plasmas because the machine is not yet built.