Reconstruction-free magnetic control of DIII-D plasma with deep reinforcement learning

Precise control of plasma shape and position is essential for stable tokamak operation and achieving commercial fusion energy. Traditional control methods rely on equilibrium reconstruction and linearized models, limiting adaptability and real-time performance. Here,the first application of deep reinforcement learning (RL) for magnetic plasma control on the mid-size DIII-D tokamak is presented, demonstrating a nonlinear approach that improves robustness and flexibility across plasma scenarios. Using the Soft Actor-Critic algorithm, this method eliminates the need for equilibrium reconstruction, enabling high-speed control execution and scalability on larger fusion devices. NSFsim, a 2D Grad-Shafranov equilibration solver with a circuit equation and a 1D transport solver, is used to train the agent. Its capability of reproducing the kinetic parameter evolution alongside magnetic equilibria evolution appears to be an essential factor significantly affecting control quality. RL-based controllers demonstrated robust magnetic control in experimental application at DIII-D, preserving control performance in transient events during plasma discharges, and reaching target parameters from the first discharge without additional tuning or modifications. The approach itself has significant generalization potential across devices and targets. This work represents a step toward AI-driven, real-time plasma control, advancing the feasibility of next-generation fusion reactors.