Physics-Guided Fast Surrogate for Real-Time Control of Vertical Stability & X-point targets in ARC-class Fusion Pilot Plant

Future fusion pilot plants like ARC[1] demands reliable, real-time control of plasma centroid and X-point positions during long-pulse (900 s) H-mode operation. Such devices will be designed to operate under limited diagnostics and without in-vessel stabilizing vertical control coils, requiring control using only superconducting out-vessel coils and central solenoid systems with stringent current and voltage limits. We present a hybrid physics-guided controller that couples a machine-learning surrogate (MEQ-ML-DN) with feedback and predictive compensation layers. MEQ-ML-DN is trained on 90k MEQ simulations [2] of ARC double-null X-point target plasma and predicts two key stability metrics: the non-rigid vertical growth rate (z) and maximum controllable/allowable displacement (Zmax) with <10% error and inference rates exceeding 5kHz. A learned control matrix maps surrogate outputs to coil commands with current-limit avoidance logic to prevent supply saturation during fast transients. Closed-loop MATLAB Simulink studies of ARC scenario shows hybrid control suppression of vertical displacements and precise X-point tracking through equilibrium evolution and forced perturbations. We report its sensitivity studies and further comparisons to ISO-FLUX boundary control policy [3].

[1] J.Helishiem & ARC team, ARC Physics Basis , APS DPP 2025, USA

[2] F. Carpanese, EPFL PhD Thesis, 2021.

[3] S. Inoue et al 2025 Nucl. Fusion 65 056020