Deep Learning for Magnetic Probes
Networks infer plasma position, shape, and current from arrays of magnetic sensors, and detect and correct faulty probe signals in real time.
Magnetics as the backbone
Magnetic pickup coils and flux loops around the vessel are the most fundamental plasma diagnostic, sensing the fields produced by the plasma current. From these signals one infers position, shape, current, and stored energy. Deep learning provides fast, robust inference from these arrays.
Inference tasks
- Plasma position and boundary shape for control
- Total plasma current and internal inductance
- Detection of magnetohydrodynamic modes from fluctuation patterns
- Identification and correction of faulty or drifting probes
A network trained on equilibria and their synthetic magnetic signatures learns to map probe readings to these quantities in a single fast pass, complementing full equilibrium reconstruction for the fastest control loops.
Fault handling
Real magnetic arrays suffer from broken channels, integrator drift, and noise. A model that has learned the redundancy among probes can flag readings inconsistent with the rest of the array and reconstruct the missing information, keeping control robust when individual sensors fail.
Mode detection
Rotating magnetohydrodynamic modes leave characteristic phase and amplitude patterns across the coil array. Learned classifiers identify the mode number and track its growth, feeding disruption warning and stability control. This turns the raw array into a real-time picture of coherent activity.
Magnetic inference underlies shape control and stability monitoring for concepts like the Hyperion breeder, including holding its -0.30 triangularity target. In Kronos design work these models are trained on simulated signatures ahead of construction; they process modeled signals because the machine is not yet built, and outputs are computational estimates.