Fusion software predicts plasma instability 200 milliseconds early

At the DIII-D National Fusion Facility in San Diego, the six microwave systems heating a fusion plasma were adjusted by software rather than a human operator. In one of the five live tests, PACMAN predicted a destructive tearing-mode instability 200 milliseconds before it could form, then changed the reactor settings early enough to prevent the disturbance altogether. Standard controllers detect these disruptions only after they have started.

PACMAN — short for Prediction And Control using MAchiNe learning — runs through four sequential stations. It first gathers temperature, density and magnetic readings, screens out measurement errors and packages the clean data. Machine-learning models then predict what the fuel will do immediately next. Controllers turn those predictions into hardware commands, such as firing a heating beam, while a final stage resolves conflicts, enforces strict equipment safety boundaries and routes the signals to the reactor.

The platform also tackled an optimization problem that had previously lacked a single working formula. It coordinated all six gyrotrons, changing their mirror angles and beam-power outputs at the same time. In other trials, a reinforcement-learning model took full control of the heating systems, while PACMAN forecast sudden bursts of energy at the edge of the fuel. The framework typically completes its work in about 20 milliseconds and then runs again and again.

In practical terms, that gives fusion researchers a control loop fast enough for events that unfold within thousandths of a second. Traditional simulations can take days or months, too slowly for experiments lasting only a few minutes; machine-learning models produce predictions in fractions of a second. PACMAN’s modular structure also means researchers can replace one algorithm in days without rebuilding the surrounding software, potentially speeding up work across other tokamak facilities.

The system is not an autonomous reactor. Physicists still set experimental goals, tune parameters between shots and use hardcoded limits to prevent equipment damage. The creators say the framework could move to other tokamaks and serve as an adaptable control standard for future commercial fusion designs, but these results remain live experimental tests rather than commercial deployment.