Bringing AI to the next generation of fusion energy

Fusion has always been the energy dream that stays five, ten, twenty years out of reach. Commonwealth Fusion Systems wants to shrink that gap with SPARC, a compact tokamak built on high-temperature superconducting magnets, and now it's getting help from Google DeepMind's AI researchers to get there faster.

The core problem is brutally hard physics: you have to hold plasma at over 100 million degrees Celsius, stable, inside a machine with real physical limits, and do it without frying the hardware. DeepMind has been chipping away at this for years, including earlier work with EPFL's Swiss Plasma Center showing that reinforcement learning could control a tokamak's magnets well enough to shape plasma on the fly. Out of that effort came TORAX, an open-source plasma simulator built in JAX that runs on both CPUs and GPUs and plays nicely with machine learning models.

With CFS, the collaboration is running on three tracks at once. First, TORAX lets CFS engineers run millions of virtual SPARC pulses before the real machine even fires up, which Devon Battaglia, CFS's senior manager of physics operations, credits with saving the team enormous amounts of setup time. Second, DeepMind is pairing TORAX with reinforcement learning and evolutionary search tools like AlphaEvolve to hunt through vast numbers of possible operating configurations and find the ones most likely to produce net energy, rather than leaving that search to trial and error. Third, and perhaps most ambitious, they're training reinforcement learning agents to manage SPARC's heat exhaust in real time, sweeping the punishing thermal load along the reactor wall in patterns too intricate for a human operator to hand-code.

Google isn't just lending researchers, either — it has also put money into CFS directly, betting on the company's path toward commercial fusion. The stated ambition goes beyond just tuning SPARC's first pulses. DeepMind talks about AI eventually sitting inside future fusion plants as an adaptive control system, not a one-off optimization tool bolted onto existing hardware. That's a much longer bet, and SPARC's actual performance once it's switched on will be the first real test of whether any of this simulation and search work translates into extra megawatts.