Google DeepMind and CFS partner to make limitless fusion energy a reality (deepmind.google)

Google DeepMind announced a research partnership with Commonwealth Fusion Systems (CFS) to accelerate delivery of net-energy fusion using AI. The collaboration pairs DeepMind’s recent advances in deep reinforcement learning and differentiable simulation with CFS’s SPARC tokamak — a compact, high‑temperature‑superconductor magnet machine targeting the historic “breakeven” milestone where fusion produces more power than it consumes. The goal: run millions of virtual experiments to identify robust operating plans, discover novel real‑time control strategies, and help SPARC reach and sustain high-performance pulses safely and sooner. Technically, the work centers on TORAX, an open‑source, JAX‑based plasma simulator that is fast, differentiable, and GPU/CPU‑friendly, enabling gradient‑based optimization and integration with learned models. DeepMind and CFS combine TORAX with reinforcement learning and evolutionary search (e.g., AlphaEvolve) to tune coil currents, fuel injection, heating, and heat‑load management (including magnetic sweeping of divertor exhaust). Prior demonstrations showed RL controlling tokamak magnetic shapes; now agents will jointly optimize power output and thermal constraints for real‑time control and pulse optimization. Calibrated against historical tokamak data and high‑fidelity codes, this stack aims to reduce commissioning time, increase operational robustness, and lay groundwork for AI-driven adaptive control in future commercial fusion plants.