AI enables faster control of unstable fusion plasma

Machine Learning Steps Into the Control Room of Fusion Plasma

Bdbusinessdaily.com – The plasma inside a tokamak is a churning, billion-degree soup of ionized gas that can tear itself apart in a fraction of a second. Human operators, no matter how skilled, simply cannot react fast enough to keep that plasma contained. Now a team of scientists at Princeton University and the Princeton Plasma Physics Laboratory (PPPL), part of the US Department of Energy’s national laboratory system, has built a software architecture that closes that reaction-time gap. The system, known as PACMAN — an acronym standing for “prediction and control using machine learning” — makes split-second decisions about how to steer fusion plasma, and it has already passed its first live-fire trials.

The researchers validated PACMAN across five separate experimental runs on the DIII-D tokamak at the National Fusion Facility in San Diego, California. Results from those runs appeared in the peer-reviewed journal Nuclear Fusion.

Why the Problem Demands Automation

Tokamaks confine plasma using intense magnetic fields shaped like a twisted doughnut. Maintaining that confinement is not a one-time setup; it demands continuous, simultaneous tweaking of heating coils, external magnets, and gas-injection valves. Instabilities — sudden disruptions in the plasma’s shape or energy distribution — can nucleate and propagate within milliseconds. A human watching a control panel simply cannot close that loop in time.

Traditional computer simulations of plasma behaviour are powerful but slow. Depending on the physics being modelled, a single simulation run can consume days or even months of compute time. That makes them invaluable for design work, yet useless as a real-time feedback mechanism during a live shot. Machine-learning models, by contrast, can ingest sensor data and project the plasma’s near-future state in a few milliseconds — fast enough to act before an instability matures.

How PACMAN Works

Rather than relying on a single algorithm, PACMAN functions as an orchestration layer that binds several distinct AI models into one coherent control pipeline. In operation, the framework continuously ingests real-time diagnostics from the tokamak: plasma temperature profiles, density measurements, and magnetic-field signals. It validates the incoming data, then routes it through the appropriate predictive models to estimate both the plasma’s present condition and its trajectory over the next few hundred milliseconds.

Armed with those predictions, internal controllers compute corrective actions — for example, ramping heating power up or down, or shifting gas-injection rates. Before any command reaches the tokamak’s actuators, PACMAN screens the proposed action against hard-coded hardware safety limits and arbitrates between models that might issue contradictory instructions. Only commands that pass every safety gate are transmitted.

What the DIII-D Trials Demonstrated

Across the five experimental runs, the system exercised a range of control functions:

A reinforcement-learning model was permitted to directly modulate plasma heating in real time. The framework forecasted bursts of energy emanating from the plasma edge. It identified and suppressed magnetohydrodynamic waves generated by fast (energetic) particles. It adjusted bulk plasma density and rotation rate on the fly. Most notably, it predicted a tearing-mode instability — a potentially device-disruptive reconnection event — roughly 200 milliseconds before the instability was expected to appear, giving operators enough lead time to reshape the plasma and prevent the event entirely rather than attempting to suppress it after onset.

In another run, PACMAN coordinated all six of DIII-D’s gyrotrons simultaneously. These devices heat the plasma with high-power microwave beams. The framework adjusted both the output power and the mirror alignment of each gyrotron in concert to hit researcher-specified heating targets, a coordination task that would be impractical to execute manually at the required speed.

Modularity and the Path Forward

One of the framework’s most practically significant features is its plug-in architecture. Integrating the first AI model into the control loop required months of engineering work. Adding a second model took only a few days. That steep drop in integration cost means researchers can test new algorithms, retrain existing ones, or swap models in and out of the loop without rebuilding the surrounding infrastructure.

“PACMAN’s modular design could allow AI algorithms to be added, replaced, or operated simultaneously without requiring changes to the rest of the system,” said Egemen Kolemen, an associate professor at Princeton University and PPPL.

The team emphasized that PACMAN is not intended to remove human judgment from the loop. Operators continue to define control objectives, set parameter boundaries, and retain final authority over experimental goals. The framework’s role is to enforce hardware safety limits unconditionally and to execute the fast-timescale adjustments that no human can perform manually.

Looking ahead, the researchers believe the architecture could be adapted to tokamaks of different geometries and scales, potentially offering a common software substrate for AI-assisted plasma control across the next generation of fusion facilities worldwide. If that portability holds, the bottleneck of manually tuning each new device’s control system could shrink dramatically, accelerating the pace at which experimental fusion programs iterate toward net-energy operation.