Inside a tokamak, a cloud of plasma hotter than the Sun must be held in place by magnetic fields that need adjusting thousands of times per second, far faster than any human can react. Since 2022, reinforcement-learning controllers from DeepMind, EPFL, and Princeton have started handling exactly that real-time balancing act on live experimental reactors. But the framing matters: these systems are decision support, not decision replacement. The AI executes the millisecond-scale control loop; physicists still design the experiments, define the target plasma shapes, and decide what “success” means.
Key takeaways
- First live control (2022): A DeepMind–EPFL reinforcement-learning agent controlled the magnetic coils of EPFL’s TCV tokamak in real time, holding and reshaping real plasma, published in Nature.
- Disruption avoidance (2024): A Princeton/PPPL team forecast tearing-mode instabilities up to 300 milliseconds ahead and steered the plasma to avoid them on the DIII-D tokamak, published in Nature.
- Humans set the goals: In both cases physicists specify target shapes and operating limits; the AI only learns how to hit them, training in simulation before touching hardware.
- Context, not hype: Fusion itself hit milestones independently of AI, including NIF’s December 2022 ignition shot (3.15 MJ out from 2.05 MJ of laser energy in).
- Industry adoption (2025): Google DeepMind partnered with Commonwealth Fusion Systems to develop an AI control “pilot” for the commercial SPARC tokamak.
Source: Degrave et al., Nature (2022)
Source: Seo, Kolemen et al., Nature (2024)
Source: LLNL / NIF (2022)
Source: ITER Organization / IAEA
Why controlling a tokamak plasma is so hard
A tokamak confines plasma at over 100 million degrees using magnetic fields generated by sets of coils. The plasma is never static: it constantly shifts, and operators must continuously adjust coil voltages to hold its position and shape. Unlike a board game with discrete moves, control is a continuous, high-dimensional problem where the right adjustments must be computed and applied in milliseconds. EPFL’s TCV uses a large set of independent poloidal-field coils, and its vertical-stability loop runs at roughly 10 kHz, the kind of speed and complexity that motivated researchers to ask whether a learned controller could do the job better than hand-tuned systems. Source: EPFL Swiss Plasma Center / EUROfusion, 2022. Read the EUROfusion summary.
2022: DeepMind and EPFL put RL on a real tokamak
In February 2022, DeepMind and EPFL’s Swiss Plasma Center reported in Nature that a deep reinforcement-learning agent had controlled real plasma on the TCV tokamak. A single neural network commanded all of the coils at once, learning directly from sensor data which voltages would produce a desired configuration. The system held conventional elongated shapes and advanced ones, including negative-triangularity and “snowflake” configurations, and even sustained a “droplet” state with two separate plasmas in the vessel simultaneously, something not previously done on TCV. Crucially, the agent was trained entirely in a simulator before being deployed, and physicists specified each target shape; the AI handled the real-time execution humans cannot. Source: Degrave et al., Nature, 2022. View the Nature paper.
2023–2024: Making it practical, and avoiding disruptions
The 2022 result was a proof of concept; follow-up work focused on making learned control reliable enough for routine use. In 2023, DeepMind and EPFL published “Towards practical reinforcement learning for tokamak magnetic control,” refining the approach to improve accuracy and robustness. Then in February 2024, a Princeton and Princeton Plasma Physics Laboratory team led by Egemen Kolemen tackled a different problem: disruptions. Working on the DIII-D National Fusion Facility in San Diego, their model, trained only on past experimental data, forecast tearing-mode instabilities up to 300 milliseconds in advance, enough time for an RL controller to adjust operating parameters and steer the plasma clear of an oncoming tear. Tearing modes are a leading cause of tokamak disruptions, so prediction-plus-avoidance is a high-value target. Source: Seo, Kolemen et al., Nature 626, 2024. View the Nature paper.
2025: AI control moves toward commercial fusion
In October 2025, Google DeepMind announced a research partnership with Commonwealth Fusion Systems (CFS) to apply AI to the SPARC tokamak. DeepMind described developing an AI “pilot” to control magnetic configurations, optimize fusion power, and manage heat load, paired with TORAX, a fast plasma simulator released in 2024 that lets AI agents explore vast numbers of operating scenarios before any are run on hardware. The move signals that learned plasma control is migrating from academic experiments toward the commercial machines that aim to deliver power to the grid in the 2030s, though SPARC has yet to demonstrate net energy. Source: Google DeepMind / Commonwealth Fusion Systems, 2025. Read the DeepMind announcement.
Where fusion stands, and what is next
AI control is one piece of a much larger fusion effort. On the inertial-confinement side, the National Ignition Facility achieved ignition on December 5, 2022, with 2.05 MJ of laser energy producing 3.15 MJ of fusion energy, the first net energy gain from a fusion reaction in a laboratory. On the magnetic-confinement side, ITER, a 35-nation tokamak in France designed for a tenfold power return (Q=10, 500 MW from 50 MW of input heating), has seen its research-operations timeline pushed to roughly 2033. None of these milestones depended on AI, but better real-time control is widely seen as essential to running the next generation of reactors steadily and safely. Source: LLNL/NIF, 2022; ITER Organization / IAEA. NIF ignition details.
The through-line: decision support, not replacement
In every result above, the division of labor is the same. Human physicists design the experiment, choose the target plasma shape, set the safety and operating limits, and judge the science. The AI does the one thing humans physically cannot: compute and apply thousands of magnetic adjustments per second to keep an unstable, ever-shifting plasma on the requested trajectory. And it does so only after training in simulation, with engineers in the loop on the real machine. Learned control replaces hand-tuned control loops, not the scientists who define the goals; it is a faster set of hands, not a new brain for fusion.
Methodology & sources
- RL agent controlled real plasma on TCV; single network drives all coils; snowflake and droplet shapes — Degrave et al., Nature (2022)
- TCV vertical-stability loop runs near 10 kHz; why continuous control is hard — EUROfusion / EPFL Swiss Plasma Center (2022)
- AI forecast tearing modes up to 300 ms ahead and avoided them on DIII-D — Seo, Kolemen et al., Nature 626 (2024)
- Refining learned control toward practical deployment — Tracey et al., “Towards practical reinforcement learning for tokamak magnetic control” (2023)
- DeepMind–CFS partnership; AI pilot and TORAX for SPARC — Google DeepMind (2025)
- NIF ignition: 3.15 MJ out from 2.05 MJ laser in, Dec 5 2022 — LLNL / NIF (2022)
- ITER: 35 nations, Q=10 / 500 MW design; revised timeline — IAEA / ITER Organization
Frequently asked questions
Does AI now run fusion reactors by itself?
No. AI controllers handle the real-time magnetic control loop, but human physicists still design the experiments, choose the target plasma shapes, and set the operating and safety limits. The systems are decision support for a control task humans cannot perform fast enough, not autonomous operators.
What did the 2022 DeepMind and EPFL result actually show?
It showed a deep reinforcement-learning agent controlling real plasma on EPFL’s TCV tokamak for the first time, reported in Nature. A single neural network drove all the magnetic coils to produce and hold a range of shapes, including snowflake and droplet configurations, after being trained in a simulator.
How does AI help prevent fusion plasma disruptions?
A 2024 Princeton and PPPL study trained a model to forecast tearing-mode instabilities up to 300 milliseconds in advance on the DIII-D tokamak. That warning gave a reinforcement-learning controller time to adjust operating parameters and steer the plasma away from a disruption before it formed.
Part of our Real-World AI Use Cases series — how AI supports high-stakes decisions across surprising domains.