DeepMind and CFS Announce Research Partnership

Last week, Google DeepMind and Commonwealth Fusion Systems (CFS) announced a formal research partnership to apply artificial intelligence to plasma simulation and control. The collaboration builds on several years of informal cooperation between the two groups and aims to improve plasma modeling and operational control. It’s worth noting that Alphabet, Google’s parent company, has been an investor in CFS since 2021 and earlier this year agreed to purchase 200 MW of electricity from the company’s first commercial plant.

CFS is developing its SPARC tokamak, planned to demonstrate net energy gain later this decade. DeepMind’s role centers on using machine-learning tools to enhance modeling accuracy and operational efficiency rather than on altering SPARC’s core design.

The joint work focuses on three areas:

A central component of the collaboration is TORAX, DeepMind’s open-source plasma simulation framework developed in JAX (Google’s high-performance numerical library). TORAX solves reduced magnetohydrodynamic (MHD) and transport equations to model the evolution of temperature, current density, and particle transport within tokamak geometries.

Its differentiable architecture allows direct coupling with machine-learning models, enabling faster optimization across a wider range of operating parameters. CFS is using TORAX to conduct large-scale virtual experiments, on the order of millions of runs, to probe SPARC’s operating space before first plasma. These simulations allow engineers to refine control scenarios, optimize plasma heating and shaping strategies, and reduce experimental iteration time once physical operations begin.

Operating a fusion tokamak requires precise control of numerous interdependent parameters, including magnet coil currents, fueling rates, auxiliary heating power, and plasma shape evolution. Given their interdependence, exhaustive manual tuning is infeasible.

DeepMind integrates reinforcement learning and evolutionary search algorithms with TORAX to automate this optimization process. The AI agents explore a wide range of operating conditions, using performance measures like confinement efficiency, plasma stability, and heat management to guide optimization. The simulations help pinpoint conditions that produce high power without exceeding design or heat-load limits.

This approach effectively functions as an autonomous optimization loop, using simulation feedback to improve control sequences before experimental validation.

The collaboration also extends to AI-based real-time control of plasma behavior. In prior work, DeepMind demonstrated a reinforcement-learning (RL) controller capable of shaping and stabilizing plasma configurations on the TCV tokamak in Lausanne, Switzerland, discovering several previously untested equilibrium trajectories.

CFS and DeepMind are now applying similar RL frameworks to SPARC’s substantially higher-power, higher-beta plasma. The system is being trained in simulation to manage dynamic magnetic equilibrium, suppress instabilities, and mitigate divertor heat flux through real-time magnetic field adjustments.

Key challenges include keeping the controller stable within millisecond-level response times and confirming that its decisions remain consistent with established plasma physics models. Hybrid control setups that combine traditional model-predictive systems with AI supervisory layers are being evaluated to improve reliability and maintain operator transparency during future experiments.

The DeepMind-CFS collaboration fits neatly within a wider trend of applying AI to fusion research. It also builds on Google’s earlier work with TAE Technologies, using machine learning to analyze plasma behavior. Machine learning is now used in diagnostics, control, and materials design across multiple fusion programs. Major computing firms have entered the field not only because of their long-term electricity needs, but also because of the natural overlap between large-scale computation and advanced plasma modeling.

While AI offers clear advantages in handling complex datasets and optimization tasks, its role in actual reactor operations is still exploratory. Control of burning-plasma devices poses safety, reliability, and interpretability challenges that remain unresolved.