Google DeepMind and Commonwealth Fusion Systems Partnership for Fusion Energy

Google DeepMind has entered a research partnership with Commonwealth Fusion Systems (CFS) to accelerate the development of viable fusion energy. The collaboration focuses on using artificial intelligence to stabilize plasma and optimize the operations of SPARC, a compact tokamak machine designed to be the first magnetic fusion device to achieve "breakeven"—generating more power from fusion than is required to sustain it.

Accelerating Plasma Simulation with TORAX

Google DeepMind is providing CFS with TORAX, an open-source, differentiable plasma simulator built in JAX. This tool allows researchers to simulate the flow of heat, electric current, and matter through the plasma core and its interactions with the surrounding systems.

Because TORAX is compatible with CPUs and GPUs and integrates with AI-powered models, it enables CFS to run millions of virtual experiments to refine operating plans before the SPARC machine is physically activated. According to Devon Battaglia, Senior Manager of Physics Operations at CFS, "TORAX is a professional, open-source plasma simulator that saved us countless hours in setting up and running our simulation environments for SPARC."

Optimizing the Path to Net Energy

Achieving maximum fusion energy requires the precise tuning of numerous variables, including heating power, fuel injection, and magnetic coil currents. To avoid the inefficiency of manual tuning, Google DeepMind is employing AI agents using reinforcement learning and evolutionary search approaches, such as AlphaEvolve.

These agents explore vast numbers of potential operating scenarios within the TORAX simulation to identify the most efficient and robust paths to generating net energy. This infrastructure allows CFS to investigate various SPARC scenarios, such as maximizing fusion power under specific constraints or optimizing for robustness, increasing the probability of success before the machine is fully commissioned.

AI-Driven Real-Time Plasma Control

Google DeepMind is developing an "AI pilot" to manage the complex real-time requirements of a tokamak. While previous work demonstrated that reinforcement learning could control magnetic configurations, the current collaboration expands this to simultaneous optimization of multiple performance metrics, including fusion power maximization and heat load management.

A critical challenge in SPARC is managing the immense heat concentrated on small areas of the machine's interior. DeepMind is investigating how reinforcement learning agents can learn to dynamically control plasma to distribute this heat effectively—for example, by magnetically sweeping exhaust energy along the wall to protect solid materials. This AI-driven approach aims to discover adaptive strategies more complex than traditional human-engineered controls and can be used to quickly tune traditional control algorithms for specific pulses.

Strategic Implications for Fusion Energy

This partnership combines Google DeepMind's AI expertise with CFS's high-temperature superconducting magnets and hardware. Beyond the immediate goal of optimizing SPARC, the vision is to establish the foundations for AI to serve as an intelligent, adaptive system at the core of future commercial fusion power plants. This technical collaboration is supported by Google's broader investment in Commonwealth Fusion Systems to move fusion technology toward commercialization.