Nuclear fusion, the process in which hydrogen nuclei collide and fuse to form heavier elements, offers the promise of vast amounts of clean energy. Achieving it on Earth requires replicating the extreme temperatures and pressures found in the core of the sun. While fusion has been realized in thermonuclear weapons, translating that reaction into a controlled, sustained process for energy generation remains one of the most formidable challenges in physics.
One of the most advanced approaches to controlled fusion uses magnetic confinement inside a tokamak. This large, torus-shaped device employs powerful magnetic fields to contain a plasma of hydrogen heated to millions of degrees. The plasma must be held stable within the magnetic field while its shape and density are constantly adjusted. Achieving this requires hundreds of precise manipulations, including targeted injections of hydrogen particles to influence plasma behavior.
Globally, only a handful of large-scale tokamaks operate, and experimental time on them is highly sought after. In the United States, the DIII-D National Fusion Facility, managed by General Atomics for the Department of Energy, is the sole machine of its kind. Researchers there explore new methods to control plasma dynamics, including the use of artificial intelligence.
DeepMind, the AI subsidiary of Alphabet, made headlines by applying reinforcement learning to control plasma confinement in the Variable Configuration Tokamak (TCV) in Lausanne, Switzerland. Their system maintained plasma stability and shaped it into various configurations, results published in *Nature* in February. Reinforcement learning, in this context, uses feedback from past and real-time experiments to determine optimal control actions.
At DIII-D, researcher Ian Char extended this concept to a different aspect of plasma behavior: its rotation. The plasma donut rotates when bursts of hydrogen particles are injected. By varying the speed and direction of these injections, it is possible to influence rotation rates and potentially improve stability. Char employed two reinforcement learning algorithms. The first was trained on years of historical tokamak data to understand how plasma responds to changes. The second monitored live plasma conditions and decided in real time how to adjust particle injection to achieve desired rotational effects.
“The short-term goal is to give the physicists the tools to cause this differential rotation so they can do the experiments to make this plasma more stable,” explained Jeff Schneider, research professor in the Robotics Institute and Char’s Ph.D. adviser. “Longer term, this work shows a path to using reinforcement learning to control other parts of the plasma state and ultimately achieve the temperatures and pressures long enough to have a power plant. That would mean limitless, clean energy for everyone.”
Char proposed the project to DIII-D and was granted a three-hour experimental window on June 28. In the facility’s control room, surrounded by operators, he loaded his algorithms into the system. The test demonstrated that reinforcement learning could indeed control plasma rotation speed—a first for the field. Some technical issues arose during the session, highlighting the need for further refinement. Char returned in late August to continue the experiments.
Egemen Kolemen, associate professor in Princeton University’s Mechanical and Aerospace Engineering Department and collaborator at the Princeton Plasma Physics Laboratory, noted, “Ian showed a tremendous ability to digest the fusion device-specific control issues and plasma physics that underlines it. It is a great achievement to apply the theory he learned at CMU to a real fusion problem and lead an experiment on a national fusion facility. That work normally requires years of plasma physics and engineering training.”
The project received support from multiple Department of Energy grants, including DE-SC0021275 for machine learning in real-time fusion plasma behavior prediction and manipulation, and DE-FC02-04ER54698. Additional funding came from the National Science Foundation Graduate Research Fellowship Program under Grant Nos. DGE1745016 and DGE2140739. The views expressed in the research do not necessarily reflect those of the NSF.