Neural-Network–Based Adaptive Control for Plasma Boundary Shaping in Tokamaks

Accurate plasma shape control in tokamaks is critical for achieving desired confinement and stability conditions, as well as for enabling other control objectives. A novel neural-network–based adaptive control strategy, which inherently incorporates real-time plasma measurements, is proposed for plasma boundary shaping. The core of this approach is a physics-informed neural network (PINN) designed to solve the inverse equilibrium problem. This problem, typically an offline optimization task that determines the necessary poloidal-field (PF) coil currents to achieve a desired plasma shape, is solved by the PINN in real-time. At each time step, the PINN receives not only the desired plasma boundary, but also real-time measurements of the plasma poloidal beta and total plasma current. Leveraging these continuous updates on the plasma's state, the PINN dynamically predicts the required PF coil currents. While the PINN's state-aware predictions provide highly accurate initial control, subtle deviations from the target shape can still arise due to inherent model approximations and unmodeled physics. To address these residual mismatches, an additional feedback controller compares the actual boundary, obtained from real-time shape reconstruction algorithms, with the desired shape and computes the needed small correction to the PF coil currents. This method offers a promising path for advanced plasma shaping.