Peering inside the machine learning black box for fusion experiments

Artificial intelligence models offer tremendous advantages in fields requiring complex simulations. Often, machine learning-based models offer better predictions of complex instabilities for a fraction of the computational resources. In fields like fusion energy, they offer real-time control of actuators to achieve more stable experimental conditions. However, their inner workings can be inscrutable.

Farre-Kaga et al. employed Shapley analysis to evaluate machine learning models used in fusion experiments. Their work can give fusion researchers insights into their models, evaluate the physical basis behind those models’ decisions, and even discover clues to new physical relationships.

The model the authors studied was designed to process inputs related to elements of the plasma profile — such as rotation, temperature, and plasma density — and output the probability of a tearing mode instability occurring. The authors’ Shapley analysis approach used statistical methods to compare the effects of different input values.

“The analysis looks at several inputs, changes the inputs, and looks at how the output is changing,” said author Hiro Josep Farre-Kaga. “So essentially, how much does this input matter to the output prediction?”

Using this approach, the authors identified core electron temperature and rotation peaking as the two most significant predictors of a tearing mode instability, with density changes having a smaller effect.

The researchers view this approach not just as a way to learn more about the models they use, but also to learn more about the physics at work in these complex environments.

“Because we know that these machine learning models perform better for predicting tearing modes, we can try to extract as much information as we can,” said Farre-Kaga. “We are trying to understand how it sees whether a tearing mode is going to happen before our physics models can.”

Source: “Interpreting AI for fusion: An application to plasma profile analysis for tearing mode stability,” by Hiro J. Farre-Kaga, Andrew Rothstein, Rohit Sonker, SangKyeun Kim, Ricardo Shousha, Minseok Kim, Keith Erickson, Jeff Schneider, and Egemen Kolemen, Physics of Plasmas (2026). The article can be accessed at https://doi.org/10.1063/5.0311201 .

This paper is part of the Papers from the 5th International Conference on Data-Driven Plasma Science Collection, learn more here .