Google DeepMind / EPFL Swiss Plasma Center · 2025 · 09 · 15 · Impact · ~2 min read

AI took control of a fusion reactor

DeepMind and Switzerland's EPFL handed plasma control of a real fusion reactor — the TCV tokamak — to a deep-learning AI. It held shapes that earlier control systems struggled with, including configurations needed for future commercial reactors. Nature, September 2025.

What's actually new

  • AI ran a real fusion plasma. Not a simulation. The TCV tokamak in Lausanne, with super-heated plasma at temperatures hotter than the sun's core, was controlled by a reinforcement-learning AI in real time.
  • Plasma shapes that were too hard before. 'Snowflake' divertors and other complex configurations needed for future power-producing reactors were stable under AI control.
  • Sim-to-real worked. The AI was trained entirely in simulation, then deployed straight onto real hardware. Earlier attempts had needed extensive on-the-machine retuning.
  • Safer ramp-down. Companion 2025 work on shutting plasma down safely (the riskiest moment) used the same approach to avoid disruptions that can damage the machine.

If you want more

Worth knowing~30s
  • 'Fusion power is closer' is a cautious claim. Better plasma control helps; commercial fusion still needs net energy gain, materials that survive years of neutron flux, and a way to produce tritium at scale. AI handles one of many problems.
  • TCV is a research reactor, not a power plant. What worked on TCV doesn't automatically transfer to ITER or commercial machines.
  • Reinforcement learning needs a good simulator. The wins here depend on physics models that aren't equally mature for every reactor type.
Who should care~20s

Anyone curious whether 'AI for science' has actual results. Energy researchers and policy people. Investors in fusion companies. Climate-focused readers who want to know whether the timeline to clean fusion is bending. Students considering a career in physics or AI for science.

What to do about it~20s

Read the DeepMind blog post (linked) — it's clearer than most science journalism on this. Don't bet on fusion power in the next decade based on this; do bet that AI plasma control will spread to other research reactors and accelerate the broader programme.

Honest take~45s

AI plasma control is the cleanest example of 'AI doing real science with hard numbers'. It doesn't make fusion easy; it makes one piece of fusion easier. The deeper story is that this same AI playbook — train in simulation, deploy on the messy real machine — is now being copied across science: weather, materials, drug discovery, and now fusion. Each individual win is modest. The cumulative effect is a quiet acceleration of how science gets done.

Other recent impact updates

Sources

Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com