A new control system based on reinforcement learning (RL) has been successfully demonstrated at the DIII-D National Fusion Facility, actively managing plasma instabilities in real time. The AI-based controller manipulates the plasma's shape and pressure profile to maintain stability, a critical step toward achieving sustained fusion reactions. This achievement, announced by the facility, represents a significant advance in applying artificial intelligence to the complex, nonlinear dynamics of magnetically confined plasmas. The system's ability to learn from experience and adapt its control strategy without a pre-programmed physics model distinguishes it from conventional control methods. Source: General Atomics / DIII-D

The DIII-D tokamak, operated by General Atomics in San Diego for the U.S. Department of Energy, is the largest magnetic fusion device in the United States. Its flexible design and advanced diagnostic capabilities make it an ideal platform for testing novel control schemes. Plasma instabilities, such as tearing modes and edge localized modes (ELMs), can lead to a rapid loss of confinement or a major disruption, potentially damaging the device and extinguishing the fusion reaction. Traditional control systems often rely on simplified models or pre-set thresholds, which can be insufficient for the dynamic and often unpredictable behavior of high-performance plasmas. Source: General Atomics / DIII-D

The DIII-D tokamak, operated by [General Atomics](/companies/general-atomics) in San Diego for the U.S.

The reinforcement learning approach differs fundamentally from these prior methods. The algorithm learns an optimal control policy through a process of trial and error, directly interacting with the plasma environment or a high-fidelity simulation. It receives a 'reward' signal for maintaining stable conditions and a 'penalty' for approaching instability boundaries. Over many iterations, the AI controller builds a sophisticated internal model of the plasma's response, enabling it to anticipate and preemptively counteract the growth of instabilities. This method allows for the control of multiple parameters simultaneously, navigating the high-dimensional operational space of a tokamak more effectively than human operators or simpler algorithms. Source: General Atomics / DIII-D

While the announcement from General Atomics did not specify the exact types of instabilities suppressed or quantify the performance improvements with metrics like disruption avoidance rates, the demonstration is a proof-of-principle for a powerful new tool. The successful application of RL at a major national facility like DIII-D validates years of theoretical and simulation-based work. The next steps will likely involve applying the technique to more challenging plasma scenarios, such as those with higher fusion power output and longer pulse durations, to test the robustness and scalability of the AI controller. Integrating such intelligent systems will be essential for the steady-state operation required by future fusion power plants. Source: General Atomics / DIII-D