Text this: Active ramp-down control and trajectory design for tokamaks with neural differential equations and reinforcement learning