Google DeepMind and Commonwealth Fusion Systems Partner to Apply AI to Next-Generation Fusion Energy

Title

Google DeepMind and Commonwealth Fusion Systems Partner to Apply AI to Next-Generation Fusion Energy

Content

Google DeepMind and Commonwealth Fusion Systems (CFS) are partnering to apply AI to SPARC, a compact high-temperature-superconducting tokamak designed to be the first magnetic fusion machine in history to generate net fusion energy. The collaboration targets three areas — fast plasma simulation, optimal-pulse discovery, and reinforcement-learning-based real-time control — and is built on TORAX, a JAX-based differentiable plasma simulator released by DeepMind last year. "TORAX is a professional, open-source plasma simulator that saved us countless hours in setting up and running our simulation environments for SPARC" — Devon Battaglia, Senior Manager of Physics Operations at CFS. Fusion, the process that powers the sun, promises clean, abundant energy without long-lived radioactive waste — but making it work on Earth means keeping an ionized gas stable at temperatures over 100 million degrees Celsius, all within a fusion machine's operating limits. The partnership builds on DeepMind's earlier collaboration with the Swiss Plasma Center at EPFL, where deep reinforcement learning successfully controlled the magnetic configuration of a tokamak to stabilize complex plasma shapes. CFS's SPARC leverages powerful high-temperature superconducting magnets and aims to be the first device to cross the so-called "breakeven" threshold — more power from fusion than it takes to sustain the plasma. The work is unfolding across three streams. First, TORAX is being deployed into CFS's daily workflows to simulate how heat, electric current and matter flow through a fusion plasma, allowing engineers to test and refine operating plans by running millions of virtual experiments before SPARC is turned on. Second, DeepMind's AI agents (using AlphaEvolve and reinforcement-learning approaches) explore vast numbers of potential operating scenarios in simulation to identify the most efficient and robust paths to net energy — adjusting magnetic coil currents, fuel injection, and heating power in concert to maximize fusion output under machine limits. Third, the team is developing an AI pilot for real-time control, extending the EPFL plasma-shape work to simultaneously optimize fusion power, manage heat load on plasma-facing materials, and balance multiple operational constraints. The platform combines TORAX's differentiability (built in JAX, so it can run on CPUs or GPUs and integrate AI surrogate models), reinforcement learning for control policies, and evolutionary search via AlphaEvolve for pulse optimization. CFS teams can use TORAX to validate against past tokamak data and high-fidelity simulations, building confidence ahead of SPARC's first operations. The DeepMind team frames the long-term ambition as "building the foundations for AI to become an intelligent, adaptive system at the very heart of future fusion power plants" — extending beyond SPARC toward the broader fusion industry.