An artificial intelligence framework tested at a US fusion research facility places a separate safety layer between machine-learning models and the equipment they control. Developed by researchers at the US Department of Energy’s Princeton Plasma Physics Laboratory and Princeton University, PACMAN (a novel abbreviation for Prediction And Control using MAchiNe learning) connects multiple AI models with a fusion experiment’s control system.

Researchers tested the framework during five experiments at the DIII-D National Fusion Facility in San Diego. According to Princeton Plasma Physics Laboratory, PACMAN collected measurements, checked the data for errors, allowed AI models to predict plasma behavior, and converted those predictions into equipment commands.

Before a command reached the equipment, a separate output stage resolved conflicts among controllers and enforced hardware safety limits.

Automated decisions occur faster than human reactions

Fusion experiments require rapid adjustments to heating systems, magnets, and gas injectors. Plasma instabilities can develop within milliseconds, making manual intervention too slow for some operating conditions.

PACMAN typically completed its control cycle in approximately 20 milliseconds. During the experiments, it controlled heating systems, adjusted plasma density and rotation, predicted energy bursts, and anticipated one type of instability approximately 200 milliseconds before it occurred.

Researchers still defined the experiment’s goals and parameters. They also reviewed each experiment before adjusting the controllers for subsequent work.

The results apply to a specialized fusion facility, and the researchers have not demonstrated that PACMAN can control routine laboratory instruments. Its architecture nevertheless provides a useful comparison with the emerging standards that connect AI agents to laboratory instruments.

Safety constraints need independent authority

An AI model can generate an incorrect command because of faulty data, an unfamiliar condition, a software change, or behavior that developers did not anticipate. Laboratories should therefore identify limits that 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 what the AI requests. Automated systems also need rules for resolving conflicting instructions and moving equipment to a safe state after a sensor, network, or software failure.

Advanced Lab Management Certificate

The Advanced Lab Management certificate is more than training—it’s a professional advantage.

Gain critical skills and IACET-approved CEUs that make a measurable difference.

PACMAN’s modular design let researchers add or replace individual models without rebuilding the entire control system. In a laboratory, that separation could support change control by allowing managers to validate a new model without redesigning established safety functions.

Human oversight also remains important when laboratories move from isolated automation to interconnected workflows. Guidance on sequencing AI-driven laboratory automation emphasizes that managers should match automation to operational readiness rather than pursue autonomy for its own sake.

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.

This article was created with the assistance of Generative AI and has undergone editorial review before publishing.