New artificial intelligence (AI) models for plasma heating can do more than was previously thought possible, not only increasing the prediction speed 10 million times while preserving accuracy but also correctly predicting plasma heating in cases where the original numerical code failed. The models will be presented on Oct. 11 at the 66th Annual Meeting of the American Physical Society Division of Plasma Physics in Atlanta.

“With our intelligence, we can train the AI to go even beyond the limitations of available numerical models,” said Álvaro Sánchez Villar, an associate research physicist at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL). Sánchez Villar is the lead author on a new peer-reviewed journal article in Nuclear Fusion about the work. It was part of a project that spanned five research institutions.

The models use machine learning, a type of AI, to try to predict the way electrons and ions in a plasma behave when ion cyclotron range of frequency (ICRF) heating is applied in fusion experiments. The models are trained on data generated by a computer code. While much of the data agreed with past results, in some extreme scenarios, the data wasn’t what they expected.

“We observed a parametric regime in which the heating profiles featured erratic spikes in rather arbitrary locations,” said Sánchez Villar. “There was nothing physical to explain those spikes.”

Sánchez Villar identified and removed problematic data, known as outliers, from the training set to train their AI since the scenarios were unphysical. “We biased our model by eliminating the spikes in the training dataset, and we were still able to predict the physics,” Sánchez Villar said. “As can be observed, the code correctly removes the spikes but anticipates higher heating in the highlighted region. However, there was nothing that would guarantee these predictions were physical,” Sánchez Villar said.

Then, the team went a step further. After months of research, the cause — a limitation of the numerical model — was identified and resolved by Sánchez Villar, who then ran the corrected version of the code for the outlier cases that were originally showing the random spikes. Not only did he find that the solutions were free of spikes in all problematic cases, but, to his surprise, these solutions were almost identical to the solutions in one of the machine learning models predicted months before, even in critical outlier scenarios (see Figure 1).
 

“This means that, practically, our surrogate implementation was equivalent to fixing the original code, just based on a careful curation of the data,” said Sánchez Villar. “As with every technology, with an intelligent use, AI can help us solve problems not only faster but better than before and overcome our own human constraints.”

As expected, the models also improved the computation times for ICRF heating. Those times fell from roughly 60 seconds to 2 microseconds, enabling faster simulations without notably impacting the accuracy. This improvement will help scientists and engineers explore the best ways to make fusion a practical power source.

Other researchers on the project include Zhe Bai, Nicola Bertelli, E. Wes Bethel, Julien Hillairet, Talita Perciano, Syun’ichi Shiraiwa, Gregory M. Wallace and John C. Wright. The work was supported by the DOE under contract number DE-AC02-09CH11466. This research used resources of the National Energy Research Scientific Computing Center (NERSC) operated under contract number DE-AC02-05CH11231, using NERSC Award FES m3716 for 2023.