Neural networks enable real-time plasma shape control in fusion reactors

Tokamak plasma control requires continuously adjusting multiple coupled shape parameters, but the mathematical tools that decouple those parameters — called virtual circuits — are too expensive to compute during an experiment. Instead of relying on a handful of pre-run reference calculations that lose accuracy as plasma drifts, this work trains neural network emulators on a library of over one million simulated Grad–Shafranov equilibria covering the MAST Upgrade operational space. Because the emulators are differentiable, their gradients yield accurate virtual circuits on demand in real time. Verification tests confirm the emulated circuits maintain high accuracy and orthogonality across a wide range of plasma configurations, making this a scalable replacement for fixed pre-scheduled control vectors in fusion devices.