Real-Time Equilibrium Reconstruction
Neural surrogates reconstruct the plasma equilibrium from magnetic measurements within the control cycle, faster than iterative reconstruction codes.
Reconstruction during operation
Feedback control of shape and position needs the plasma equilibrium many times per second. Classical reconstruction fits magnetic measurements to a Grad-Shafranov solution iteratively, which can be too slow for the fastest control loops. A learned surrogate maps measurements to the equilibrium in a single fast pass.
Inputs and outputs
- Inputs: magnetic probe and flux-loop signals, coil currents
- Outputs: flux map, boundary shape, position, and derived quantities
- Optional: internal profile constraints from other diagnostics
The network is trained on a large set of reconstructions from a classical code, learning to reproduce them at a fraction of the time. Because magnetic signals underdetermine the internal current profile, the surrogate learns the same regularizing assumptions the reference code uses.
Real-time constraints
Deployed reconstruction runs on deterministic real-time hardware with a hard latency budget. A neural surrogate's fixed-cost forward pass fits this requirement better than an iterative solver whose convergence time varies. Predictable timing is as important as raw speed.
Accuracy and fallback
The surrogate must stay accurate across the full operating range and flag when a measurement set is anomalous, for example after a sensor fault. Pairing it with uncertainty estimates and a slower reference reconstruction for periodic cross-checks guards against silent errors during control.
Fast reconstruction underpins holding a target shape, such as the -0.30 triangularity of the Hyperion breeder concept. In Kronos design work these surrogates are trained and tested on simulated equilibria before construction; they process modeled signals because the device is not yet built, and outputs are computational estimates.