New AI model of plasma heating leads to important modifications to computer code used in fusion research

A new artificial intelligence (AI) model for plasma heating can do more than previously thought possible, increasing prediction speed by a factor of 10 million while maintaining accuracy, as well as Plasma heating can be accurately predicted even if the code fails. These models will be presented at the 66th American Physical Society Plasma Physics Division Annual Meeting on October 11 in Atlanta.

“With our intelligence, we can push AI even further beyond the limits of available numerical models,” said Alvaro Sánchez Villar, associate research physicist at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL). It can be trained.” Sánchez Villar is the lead author of a new journal article in the journal Fusion about this research. This was part of a project spanning five research institutions.

This model uses machine learning, a type of AI, to predict how electrons and ions in a plasma will behave when ion cyclotron frequency range (ICRF) heating is applied in a fusion experiment. I’ll try. The model is trained on data generated by computer code. Although much of the data was consistent with past results, in some extreme scenarios the data was not as expected.

“We observed a parametric situation where the heating profile was characterized by irregular spikes at fairly arbitrary locations,” Sánchez-Vilal said. “There was nothing physical to explain those spikes.”

Sánchez Villar identified and removed problematic data, known as outliers, from the training set for training the AI ​​because the scenario was non-physical. “We biased the model by eliminating spikes in the training dataset, and it was still able to predict physics,” Sánchez-Vilal said.

“As you can see, the code is accurately removing the spikes, but predicting higher heating in the highlighted areas. But what guarantees that these predictions are physical? There wasn’t.”

Then the team went one step further. After several months of research, the cause (a limitation of the numerical model) was identified and resolved by Sánchez-Vilal. Sánchez-Bilal then ran the modified version of the code for outlier cases that originally showed random spikes.

Not only did he find that there were no spikes in his solutions in all problematic cases, but surprisingly, these solutions predicted months in advance, even in severe outlier scenarios. The solution for one of the machine learning models was similar.

“This effectively means that, based on careful collection of data, our surrogate implementation is equivalent to modifying the original code,” Sánchez-Villar said. “As with any technology, when used wisely, AI can not only solve problems faster, but better than before, and overcome our own limitations.”

As expected, this model also reduced the calculation time for ICRF heating. These times were reduced from approximately 60 seconds to 2 microseconds, allowing for faster simulations without significantly impacting accuracy. This improvement will help scientists and engineers explore the best ways to make fusion a practical power source.