Scientists have built a machine learning system that can predict tiny shifts in the hardware of a major fusion research device before the next experiment begins. The framework, developed by data scientists at the U.S. Department of Energy’s Thomas Jefferson National Accelerator Facility and colleagues, focuses on the DIII-D National Fusion Facility in San Diego and could help keep plasma experiments more stable.
DIII-D is a tokamak, a doughnut-shaped machine that holds plasma hotter than the sun’s core inside a powerful magnetic field. The field is produced by a ring of large magnets known as toroidal field coils. Even though the coils are bolted firmly in place, they can move slightly from one shot to the next as the plasma changes. Those small movements can affect the physics inside the machine.
Shots at DIII-D occur roughly every 10 minutes. Researchers wanted a way to forecast coil movement during the short break between experiments so operators could adjust plasma settings or schedule maintenance if needed.
“The technique that we have developed would help fusion researchers find issues before they actually happen on the physical machine,” said Kishan Rajput, a data scientist at Jefferson Lab and lead author of the study. “This takes us one step closer to putting AI in operations for fusion diagnostics.”
The team created a digital twin of the coil system. It uses deep neural networks trained through online learning, meaning the models continuously update as new data arrives. Because plasma behavior drifts from shot to shot, a model trained only on old data would quickly lose accuracy.
Multiple models trained on different time windows form an ensemble. Some track sudden changes, others gradual ones. Each prediction comes with an uncertainty measure. The system weights the more confident models more heavily when combining results.
Compared with static machine learning models, the online approach cut prediction error by 80 percent. Adding the uncertainty-guided ensemble trimmed error by another 10 percent relative to a single online model.
“Unexpected movement in the physical machine would not be desired,” Rajput said. “You want to predict the movement of these coils during experiments to get a sense of how stable a particular shot would be and whether a problem may arise.”
“We are creating a virtual replica, more or less,” he added. “You can feed all sorts of different parameters to this virtual replica that you may want to run on the physical machine.”
The work was published in the journal Machine Learning with Applications. Partners included General Atomics, the University of Houston and the Pacific Northwest National Laboratory.
Rajput plans further tests on years of historical data to capture rarer events and improve uncertainty estimates. He also wants the models to reveal more clearly how they are evolving, which would help build trust in the system.
“People want to know what’s going on inside an AI model as opposed to it being a so-called ‘black box,’” he said. “That’s a top priority because it builds trust.”
The framework is ready for use at DIII-D and could be adapted for other fusion machines. Researchers say the combination of adaptive learning, uncertainty measures and real-time constraints makes it practical for actual operations, a step forward for the field.
