Fusion Lab Tests AI Safety Framework That Keeps Hardware Limits Outside the Model

Fusion Lab Tests AI Safety Framework That Keeps Hardware Limits Outside the Model

Researchers at Princeton Plasma Physics Laboratory and Princeton University have tested an artificial intelligence control framework that places hardware safety constraints in a separate layer outside the machine-learning models themselves. The system, called PACMAN (Prediction And Control using MAchiNe learning), was validated during five experiments at the DIII-D National Fusion Facility in San Diego.

According to Princeton Plasma Physics Laboratory, PACMAN connects multiple AI models to a fusion experiment's control system through a modular architecture. The framework collects sensor measurements, validates data quality, allows AI models to predict plasma behavior, and converts those predictions into equipment commands. Critically, before any command reaches heating systems, magnets, or gas injectors, a separate output stage resolves conflicts among controllers and enforces predefined hardware safety limits.

Why it matters

As laboratories move from using AI to analyze data toward letting it control physical equipment, the question of how to constrain automated decisions becomes urgent. PACMAN's architecture demonstrates that safety limits need not be embedded within—and potentially overridden by—the AI model itself. Instead, independent constraint layers can act as a final authority regardless of what the model requests, a design principle that applies beyond fusion research to any laboratory automation where incorrect commands could damage equipment or compromise safety.

Millisecond decisions outpace human reaction

Fusion experiments require rapid adjustments because plasma instabilities can develop within milliseconds. PACMAN typically completed its control cycle in approximately 20 milliseconds during the experiments. The system controlled heating systems, adjusted plasma density and rotation, predicted energy bursts, and anticipated one type of instability roughly 200 milliseconds before it occurred—faster than manual intervention allows.

Researchers still defined experiment goals and parameters and reviewed each run before adjusting controllers for subsequent work. Human oversight remained central even as automated decisions accelerated.

Modular design supports validation and change control

PACMAN's modular structure allowed researchers to add or replace individual AI models without rebuilding the entire control system. In laboratory settings, this separation could support change control by enabling managers to validate a new model without redesigning established safety functions.

The researchers have not demonstrated that PACMAN can control routine laboratory instruments, and the results apply specifically to a specialized fusion facility. Nevertheless, the architecture offers a useful comparison point as emerging standards connect AI agents to laboratory equipment.

Independent constraints for laboratory automation

An AI model can generate incorrect commands due to faulty data, unfamiliar conditions, software changes, or unanticipated behavior. Laboratories should therefore identify limits the model cannot override, including maximum temperatures, pressures, speeds, forces, volumes, and travel distances.

Emergency stops, guards, door interlocks, collision detection, and containment controls should remain effective regardless of AI requests. Automated systems also need rules for resolving conflicting instructions and moving equipment to safe states after sensor, network, or software failures.

As AI moves from interpreting data to controlling physical equipment, lab managers must evaluate more than model accuracy. They need to determine which actions the system can take, which limits remain fixed, how commands are recorded, and when a human must approve or interrupt the work.

These details were first reported by Lab Manager.