AI Takes Control of Fusion Systems in Milliseconds

Fusion researchers are trying to control matter at temperatures hotter than the center of the Sun. That challenge leaves almost no room for delayed decisions. A plasma instability can develop in milliseconds, while a detailed physics simulation may require days or months to complete. Human operators can set the goals, but they cannot manually react quickly enough to every change inside a fusion experiment.

A new study¹ from researchers at Princeton University and the Princeton Plasma Physics Laboratory describes a practical response to that control problem. The team developed PACMAN, short for Prediction And Control using MAchiNe learning, and deployed it on the DIII-D tokamak in California. Instead of performing one narrow AI task, PACMAN provides a common framework through which different models can read diagnostic data, make predictions, coordinate controllers and send commands to physical equipment.

Fusion research already has powerful machine learning models. The harder problem is turning them into dependable real-time components without rebuilding the control pipeline for every experiment. PACMAN addresses that missing layer.

Why Fusion Requires Real-Time AI Control

A tokamak confines electrically charged plasma using magnetic fields. Maintaining useful conditions requires constant adjustments to heating systems, magnetic coils, gas injection and other equipment. Small disturbances can grow into instabilities that reduce performance or terminate the plasma discharge entirely.

Traditional controllers remain effective when engineers can describe the desired response in advance. However, advanced plasma regimes involve nonlinear interactions, incomplete measurements and conditions that change too quickly for exhaustive calculations. Machine learning can approximate this behavior and deliver an answer within the available time.

The broader importance of this approach has already become visible. Previous research showed that AI can predict and prevent fusion instabilities before conventional systems would respond. PACMAN extends the concept from an individual prediction and control demonstration into an architecture capable of supporting several models, controllers and objectives.

According to the accompanying PPPL announcement, the complete framework typically runs in approximately 20 milliseconds. In one experiment, a machine learning model anticipated a tearing-mode instability about 200 milliseconds in advance. That interval may appear brief, but it gives an automated controller enough time to redirect heating and change plasma conditions before the instability becomes established.

How The PACMAN Fusion Framework Works

PACMAN organizes real-time control into four consecutive blocks. The input block gathers diagnostic measurements and checks their validity. The model block allows machine learning, conventional or physics-based models to interpret those inputs. The controller block translates predictions into desired actions. Finally, the output block resolves competing requests, applies hardware constraints and passes approved commands to the tokamak.

The system can therefore separate intelligence from authority. An AI model may recommend an action, but its output does not bypass the surrounding control architecture. Commands are still checked against actuator limits, and researchers define the objectives and configuration for each experiment.

The framework includes several practical safeguards:

• Input validation identifies missing or invalid diagnostic data.
• Low-pass filtering can prevent reactions to isolated poor measurements.
• Conflict handling arbitrates between competing actuator requests.
• Output limits prevent commands from exceeding hardware constraints.

This design offers a useful lesson for safety-critical AI: reliability does not have to depend on an infallible model. The surrounding system can assume that models will encounter bad data or produce unsuitable outputs, then constrain their influence over physical equipment.

Five Experiments Show The Framework’s Flexibility

The researchers used PACMAN in five experimental applications on DIII-D. These were not five variations of the same neural network. They included reinforcement learning, survival models, reservoir computing and model-predictive control, paired with different diagnostics and actuators.

The reinforcement-learning controller managed heating inputs to target plasma pressure and internal transport barriers. The tearing-mode controller redirected electron cyclotron heating before an instability appeared. Another model predicted sudden edge-localized events without directly controlling an actuator, showing that PACMAN can also supply an external controller.

The Alfvén eigenmode application classified waves associated with energetic particles and adjusted neutral-beam power in response. The most computationally demanding example compressed plasma profiles into a simpler latent representation and used model-predictive control to adjust density and rotation toward researcher-defined targets.

Together, these experiments show why a shared framework is valuable. Each objective requires different data, timing and hardware, but every team should not need to recreate diagnostic handling, safety checks and actuator integration.

The Real Breakthrough Is Faster Fusion Development

PACMAN does not solve the scientific and economic barriers facing fusion power. DIII-D is a research facility, not a commercial generator, and the experiments did not demonstrate net electricity. The framework also cannot control phenomena that evolve faster than its millisecond-scale window. Its current implementation executes models serially, although the architecture can support future parallel execution.

Its near-term value is instead found in development speed. PPPL reports that building the framework and installing the first model took months, while adding the second model required only days. Once shared infrastructure handles recurring integration problems, researchers can spend more time improving plasma models and less time rebuilding plumbing around them.

This could create a compounding effect. Faster deployment produces more experimental data, better models and increasingly ambitious operating regimes. The result is a tighter learning loop between simulation, machine learning and physical experiments.

That loop helps explain the growing connection between AI and fusion energy. AI may increase electricity demand through data centers, but it is also becoming an engineering tool for developing new energy systems. In fusion, its value is not limited to discovering materials or analyzing results after an experiment. It can participate directly in controlling the machine.

From Experimental Software To Reactor Infrastructure

The next step will be determining whether PACMAN’s design principles transfer to other machines. Tokamaks differ in their geometry, diagnostics, actuators and control systems. A reusable conceptual architecture does not automatically make software portable across every facility.

Even so, the framework identifies elements likely to matter in future reactors: standardized inputs, interchangeable models, explicit controller boundaries, actuator arbitration and safety enforcement outside the AI model. Commercial systems will also require redundancy, cybersecurity and predictable responses when equipment fails.

The researchers plan to introduce multithreading and explore GPUs or field-programmable gate arrays for faster inference. The critical metric will not be raw model speed alone. End-to-end latency, including data acquisition, validation, decisions and actuation, determines whether a controller can intervene in time.

Investing In AI-Enabled Fusion Through Alphabet

For investors seeking public-market exposure to the convergence of AI and fusion research, Alphabet Inc (GOOG ) offers a relevant, although indirect, option. Google has collaborated with private fusion developer TAE Technologies for more than a decade, applying machine learning to plasma reconstruction, optimization and reactor design. Google also participated in TAE’s 2025 financing round.

Google’s contribution extends beyond capital. Its researchers have worked with fusion engineers to interpret plasma behavior and reduce machine complexity. In 2025, Google said joint plasma reconstruction work helped TAE simplify and rebuild its fusion machine. Alphabet has also agreed to purchase power from Commonwealth Fusion Systems if its planned facility reaches commercial operation.

This does not make Alphabet a fusion pure play. Fusion currently represents a small and speculative part of a company dominated by advertising, cloud computing and digital services. TAE also uses a field-reversed configuration rather than the tokamak architecture employed at DIII-D. Still, Google’s AI collaboration with TAE Technologies demonstrates that it is contributing technical capabilities to the sector rather than waiting to become a future electricity customer.

AI Is Becoming Part Of The Fusion Machine

PACMAN moves the discussion beyond whether one AI model can predict an instability. The more consequential question is whether many predictive and control systems can operate together, interact with hardware and remain bounded by enforceable safety rules.

The DIII-D experiments suggest that this integration problem is solvable. PACMAN accommodated five distinct applications, made decisions on millisecond timescales and provided a path for adding new models without redesigning the entire control system. It remains experimental infrastructure, but infrastructure is precisely what turns isolated breakthroughs into repeatable progress.

Commercial fusion still requires advances in plasma performance, materials, fuel cycles, maintenance and plant economics. Real-time AI control will not replace those achievements. It may, however, become the layer that allows them to work together inside a reactor.

References:

^{1 Rothstein, A., Farre-Kaga, H. J., Butt, J., Shousha, R., Erickson, K., Wakatsuki, T., Steiner, P., Kim, S. K., Jalalvand, A., & Kolemen, E. (2026). Enabling integrated AI control on DIII-D: A control system design with state-of-the-art experiments. Nuclear Fusion, 66(7), 076050. https://doi.org/10.1088/1741-4326/ae7f9d}