AI Framework PACMAN Achieves Millisecond-Level Control of Fusion Plasma

Researchers at Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) have handed split-second control of a live fusion reactor to an artificial intelligence system, and are now preparing to scale the software to larger experimental facilities. The framework, known as PACMAN, successfully predicted and prevented a plasma disruption 200 milliseconds before it could form during tests at the DIII-D National Fusion Facility in San Diego, proving that machine learning models can manage the extreme conditions of magnetic confinement fusion at speeds human operators cannot match.[1][2][4]

Inside a tokamak, plasma must remain extraordinarily hot and stable to sustain a fusion reaction. However, particles hotter than the core of the sun can become unruly in a few thousandths of a second. Small disturbances can grow in milliseconds, creating instabilities that disrupt the plasma far faster than conventional simulation software can process.[1][2]

To address this timing problem, the Princeton team developed PACMAN—short for Prediction And Control using MAchiNe learning. The system operates as a modular, four-stage control loop. It continuously ingests real-time tokamak telemetry, including temperature, density, and magnetic signals. It then validates the data and routes it to specialized machine learning models for behavior prediction.[1][3][4]

The speed of the system represents a fundamental shift in how fusion experiments are managed. "A really focused human operator can respond on the order of seconds," said Andy Rothstein, a graduate student at Princeton University's Department of Mechanical and Aerospace Engineering. "The whole PACMAN framework typically runs in about 20 milliseconds, and it's not running once. It's running again and again and again. It can see small things happening in the plasma and adjust in a way that a human would never be able to do."[1][3]

During the September 2026 experimental campaigns, which build on architectural designs first outlined in November 2025, the Princeton team deployed the framework on the DIII-D tokamak to execute five separate plasma-control tasks. Rather than relying on a single AI model, PACMAN integrates multiple machine learning models and controllers into the same loop. An output stage resolves conflicting directives and enforces strict hardware safety limits before transmitting adjustments to the reactor.[4][5]

One of the most significant tests involved preventing a tearing-mode instability, a phenomenon that disrupts the magnetic structure confining the plasma. Conventional controllers generally respond after the instability has already started, which often leads to performance degradation as the system struggles to suppress the disturbance.[2][3]

PACMAN took a preventative approach. "In one of the experiments we present, a machine learning model predicts the tearing mode about 200 milliseconds in advance, so the plasma can be changed to avoid it in the first place," said co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics.[1][3]

The framework also simultaneously steered all six of DIII-D's gyrotrons—systems that heat the plasma with powerful microwave beams—to achieve complex thermal targets established by researchers before the experiment began. PACMAN adjusted the gyrotrons' power while repositioning their mirrors in real time. "There was no algorithm to find that optimal solution before," Farre Kaga noted.[1][3]

Because PACMAN relies on standardized data interfaces and independent algorithmic modules, individual AI components can be added, updated, or replaced without disrupting the broader system. Researchers noted a dramatic reduction in development time for subsequent models, with iteration cycles shrinking from months to days. The system's design and preliminary experimental results were detailed in the journal Nuclear Fusion.[4][5]

By compressing the control timeline to the millisecond scale while preserving human-in-the-loop safety protocols, PACMAN establishes a new operational standard for magnetic confinement fusion. The framework provides a reusable software foundation that the research team plans to expand from DIII-D to tokamak devices of different sizes and configurations, removing a major technical hurdle on the path to commercially viable fusion energy.[1][4]