Bringing AI to the next generation of fusion energy

Source: DeepMind Date: 2025-10-23 URL: https://deepmind.google/blog/bringing-ai-to-the-next-generation-of-fusion-energy/

Summary

Google DeepMind partnered with Commonwealth Fusion Systems to apply deep RL to tokamak plasma control for CFS’s SPARC reactor. Core deliverable is TORAX — an open-source, JAX-based differentiable plasma simulator now integrated into CFS’s daily workflows — enabling millions of virtual experiments before SPARC is operational. Prior work demonstrated RL controlling tokamak magnets to stabilize complex plasma shapes; this extends that to full operating scenario optimization for net energy generation (breakeven).

Implications

TORAX as simulation infrastructure, not just a research demo. Open-sourcing a differentiable plasma simulator that CFS uses daily is how DeepMind builds lasting positioning in fusion energy R&D — not by keeping tools proprietary but by becoming the simulation infrastructure layer for the field. That’s the AlphaFold playbook applied to plasma physics.

RL-plus-search in the physical sciences thread. LIGO noise reduction (Deep Loop Shaping), tokamak plasma control, gravitational wave detection — DeepMind is systematically applying the same RL + domain physics methodology across precision science applications. The pattern is: where traditional control theory hits limits, RL fills the gap.

Fusion energy is long-horizon but CFS is the most credible near-term bet. SPARC is targeting breakeven in the late 2020s. If AI-optimized plasma control improves the probability of achieving that from day one of operations, the impact is strategic — not just academic. Google getting embedded in CFS’s workflow now positions them well for that milestone.

Watch:

• TORAX adoption by other fusion research programs (MIT PSFC, ITER, TAE Technologies)
• Whether SPARC achieves breakeven and whether AI control is cited as a contributing factor
• DeepMind’s fusion science publications using TORAX simulation results