Machine learning takes command of fusion plasma as SPARC approaches first light

DeepMind’s AI will control the magnetic fields confining SPARC’s fusion plasma, learning optimal strategies through millions of virtual experiments before the tokamak’s first firing

(Image courtesy of Commonwealth Fusion Systems)

The race toward commercial fusion energy has entered a new computational frontier. Commonwealth Fusion Systems is deploying Google DeepMind’s artificial intelligence infrastructure across its SPARC tokamak project in Massachusetts, fundamentally changing how engineers will attempt to control the superheated plasma that must be sustained at over 100 million degrees Celsius for fusion reactions to occur.

The partnership centers on TORAX, an open source plasma transport simulator built in the JAX framework, combined with reinforcement learning systems that can explore millions of operational scenarios before the machine ever fires. For an industry historically bound by manual experimentation and conservative operational envelopes, this represents a substantial departure from established practice.

The control problem

Containing fusion plasma presents a multidimensional control challenge that exceeds human reaction time. Inside a tokamak, charged particles heated to temperatures far beyond the sun’s core must be held away from the vessel walls by precisely shaped magnetic fields. The plasma itself is inherently unstable. Small perturbations in density, temperature, or magnetic field geometry can cascade into disruptions that terminate the fusion reaction in milliseconds.

Traditional control architectures divide this problem into separate subsystems. Each of SPARC’s magnetic coils typically operates under its own proportional-integral-derivative controller, with additional layers managing plasma shape, position, and current profile. These systems must coordinate thousands of adjustments per second while respecting hard physical limits on coil voltages, thermal loads, and structural forces.

DeepMind’s work at the Swiss Plasma Center’s TCV tokamak in Lausanne demonstrated that a single neural network could replace this distributed control architecture entirely. Published in Nature in 2022, the research showed that deep reinforcement learning could command all 19 of TCV’s magnetic coils simultaneously, learning optimal voltage patterns directly from sensor data. The system successfully maintained elongated plasma configurations with vertical instability growth rates of 1.4 kHz and produced novel geometries including sustained plasma droplets, where two separate plasma regions existed simultaneously within the vacuum vessel.

TORAX architecture and physics

TORAX solves coupled one-dimensional partial differential equations for ion heat transport, electron heat transport, particle transport, and current diffusion across the plasma core. The simulator uses a finite volume method to discretize these governing equations spatially, tracking how temperature, density, and current profiles evolve through time.

The framework incorporates multiple physics models. For turbulent transport, which determines how quickly heat and particles diffuse outward from the plasma core, TORAX couples to QLKNN neural network surrogates. These machine learning models were trained on outputs from QuaLiKiz, a gyrokinetic code that captures the complex electromagnetic turbulence driving transport in magnetized plasmas. The QLKNN models accept local plasma parameters as inputs, including normalized temperature and density gradients, safety factor, magnetic shear, effective charge, and collisionality, then output turbulent flux predictions for heat and particle channels.

QLKNN achieves computational speeds several orders of magnitude faster than direct gyrokinetic simulations while maintaining accuracy within the model’s validated regime. The latest QLKNN_7_11 version includes improved collision operators and better coverage of near-edge regions compared to earlier iterations. TORAX implements optional Gaussian convolution smoothing of these transport coefficients to improve solver convergence when dealing with particularly stiff transport models.

For neoclassical physics, TORAX uses the analytical Sauter model to calculate bootstrap current and neoclassical conductivity. The simulator also accounts for fusion power generation, ohmic heating from resistive dissipation of plasma current, ion-electron heat exchange, Bremsstrahlung radiation losses, and impurity radiation. An ion cyclotron resonance heating surrogate model provides auxiliary heating capabilities, though this currently covers limited operational regimes.

Built in JAX, TORAX compiles to run efficiently on both CPUs and GPUs with full support for automatic differentiation. This differentiability enables gradient-based nonlinear PDE solvers, trajectory optimization, and sensitivity analysis of simulation outputs to arbitrary input parameters. Researchers can analytically compute how changes in heating power, fueling rate, or magnetic field configuration propagate through the entire simulation without hand-deriving Jacobians.

The code was verified against RAPTOR, an established tokamak transport simulator, demonstrating agreement in predicted plasma profiles across multiple test cases. TORAX now operates as CFS’s primary tool for daily physics workflow, replacing what had previously been a fragmented collection of simulation packages written in different languages.

Reinforcement learning for pulse optimization

The collaboration extends beyond fast simulation into autonomous control strategy development. CFS and DeepMind are investigating how reinforcement learning agents can learn to dynamically manage SPARC operations by running vast numbers of virtual experiments in TORAX.

The approach treats tokamak operation as a sequential decision problem. At each time step, the control system observes the current plasma state from diagnostics and must select control actions, including adjustments to magnetic coil currents, neutral beam or radio-frequency heating power, and fuel injection rates. The system receives reward signals based on how well it achieves operational objectives while respecting safety constraints.

During training, the reinforcement learning agent explores different control sequences in simulation, gradually learning which actions lead to better outcomes. The training process incorporates domain randomization, varying plasma physics parameters and initial conditions across simulated shots to build robust policies that can handle uncertainty and measurement noise.

For SPARC specifically, one critical application focuses on heat exhaust management. Fusion reactions deposit tremendous power that must be safely removed from the plasma. This exhaust concentrates in a region called the divertor, where plasma-facing materials encounter extreme heat fluxes. By precisely controlling plasma shape and magnetic field geometry, operators can distribute this thermal load more evenly across the divertor target plates.

The reinforcement learning system can learn control policies that actively shape the plasma to manage divertor heat loads while simultaneously optimizing fusion power output. This represents a fundamentally different paradigm than conventional control, where engineers specify explicit trajectories and then implement tracking controllers to follow them. Instead, the AI system discovers control strategies by directly optimizing for specified objectives.

DeepMind has also developed AlphaEvolve, an evolutionary search approach that complements pure reinforcement learning. These algorithms can explore the space of possible SPARC pulse scenarios to identify configurations that maximize fusion power under various operational constraints or optimize for operational robustness as understanding of the machine improves.

Integration with digital twin development

SPARC’s AI-driven control development runs parallel to an ambitious digital twin program being developed with NVIDIA and Siemens. The digital twin synthesizes mechanical, thermal, electromagnetic, and plasma physics simulations into a unified virtual representation of the entire facility.

NVIDIA’s Omniverse platform provides the integration framework, using OpenUSD standards to combine data from classical physics solvers with AI surrogate models. Siemens contributes Teamcenter for product lifecycle management and Designcenter NX for computer-aided design and mechanical analysis. SPARC comprises over 2 million individual components, comparable in complexity to a commercial aircraft, and many components shift position during operation as magnets contract when cooled and the vacuum vessel expands when heated.

The digital twin enables physicists in the control room to compare real-time experimental data against simulation predictions during actual shots. When SPARC fires, operators can immediately see how measured plasma profiles, heating efficiency, and confinement performance compare to what the models predicted. This tight feedback loop accelerates the learning process and helps identify when the machine operates outside validated model regimes.

Traditional plasma transport simulations might require weeks of computing time on conventional architectures. NVIDIA’s AI infrastructure enables training of surrogate models that capture complex plasma behavior while running millions of times faster than first-principles codes. These neural network surrogates learn to approximate the computationally expensive physics calculations, transforming what would be month-long simulation campaigns into millisecond inferences that can execute in real time.

The digital twin also serves scenario planning before machine operation begins. Engineers can test proposed pulse sequences, evaluate component thermal and mechanical loads, and identify potential operational issues entirely in simulation. This front-loads risk reduction and allows the team to focus experimental time on the most promising operating regimes once SPARC starts commissioning.

Technical challenges and machine learning limitations

Despite the sophisticated AI infrastructure, significant technical hurdles remain. TORAX currently operates on a uniform spatial grid and lacks detailed models for auxiliary heating and current drive systems beyond basic formulations. Future development aims to incorporate more sophisticated physics-based models or additional ML surrogates for neutral beam injection, ion cyclotron heating, and electron cyclotron current drive.

The transport models themselves, particularly the QLKNN neural networks, have defined validity ranges based on their training data. Operating outside these regimes reduces prediction accuracy and potentially introduces systematic errors in optimization results. TORAX implements checks to flag when simulations venture beyond validated parameter space, but this inherently limits the scope of autonomous exploration.

Reinforcement learning controllers trained in simulation face the fundamental sim-to-real transfer problem. Even with domain randomization and high-fidelity models, discrepancies between simulated and actual tokamak behavior can cause learned policies to underperform or fail entirely when deployed on hardware. The TCV experiments mitigated this by incorporating detailed power supply dynamics, sensing models, and realistic environmental parameter variation during training, but SPARC represents a substantially larger and more complex machine operating in an entirely unexplored region of parameter space.

Machine protection remains paramount. While AI systems can learn sophisticated control strategies, they must operate within hard constraints on coil voltages, structural loads, and plasma operating limits. CFS’s implementation will likely retain conventional safety systems as fallback controllers and interlocks, ready to assume control if the AI system attempts an unsafe action or encounters an unexpected scenario.

Implications for fusion development timeline

CFS aims to operate SPARC and demonstrate net energy gain, where fusion power output exceeds the power required to sustain the reaction, within the next few years. The company’s subsequent ARC power plant is projected to begin delivering 200 megawatts to Google’s data centres in Virginia in the early 2030s.

These timelines are aggressive by fusion energy standards. Applying AI throughout the development process represents an attempt to compress what would traditionally require decades of iterative experimental campaigns into a substantially shorter development cycle. The technology enables testing millions of operational scenarios virtually before committing precious experimental time, potentially identifying optimal operating regimes that would take years to discover through manual exploration.

However, fusion energy has repeatedly proven more difficult than anticipated. SPARC itself relies on high-temperature superconducting magnets that have only recently been validated at the required scale. The interaction between advanced magnet technology, AI-optimized plasma control, and novel operational regimes introduces layers of complexity and potential failure modes that are difficult to predict even with sophisticated simulation.

The broader fusion community is watching the CFS-DeepMind partnership closely. If successful, the approach could establish a new paradigm where machine learning becomes integral to fusion reactor development from initial design through commercial operation. The techniques being developed for SPARC are intentionally general and extensible. TORAX is open source, and the reinforcement learning methods could potentially transfer to other fusion concepts beyond tokamaks.

What remains to be demonstrated is whether AI-driven optimization can actually accelerate the path to commercial fusion energy or whether the fundamental physics challenges will prove resistant to even the most sophisticated computational approaches. SPARC’s first plasma shots will provide the first experimental data to validate or refute the simulation-optimized strategies that AI systems have proposed.

The stakes extend beyond fusion itself. As Google and other technology companies face exponentially growing energy demands from AI data centre, securing clean, baseload power has become strategically critical. The intersection of machine learning driving both fusion development and the energy demand motivating fusion investment creates an unusual feedback loop in energy technology development.

For now, SPARC assembly continues at the Devens facility. The tokamak’s magnet systems are being installed, vacuum vessel components are being integrated, and control systems are being commissioned. In parallel, the TORAX simulations run continuously, exploring the operational space that will soon become experimental reality. Whether the AI systems can deliver on their promise remains an open question, but the fusion community has never before had computational tools of this sophistication applied to the control problem at this scale.