Scientists from the U.S. Department of Energy's (DOE) Princeton Plasma Physics Laboratory (PPPL) use machine learning (ML) to create a model for rapid control of plasma. Plasma is the phase of matter comrpised of atomic nuclei and free electrons, or ions that power fusion reactions.

The researchers target to use machine learning to understand clean fusion energy. Face recognition, language use, self-navigating cars are all within the realm of artificial intelligence.

Plasma makes up the sun and most stars. These undergo constant fusion reactions. Scientists have to find a method on how to fuse and release the energy of plasma. The researchers believe ML can control such.

Neural Networks

Neural networks, the heart of ML software, are trained in producing data in the first operational campaign of the flagship fusion facility, National Spherical Torus Experiment-Upgrade (NSTX-U), at PPPL. The behavior of the energetic particles produced by powerful neutral beam injection (NBI) is accurately predicted by this trained model. These particles power NSTX-U plasmas and heat them to million-degree, fusion-relevant temperatures.

NUBEAM, a complex computer code, generates these predictions. This program integrates data regarding the effect on the plasma by the beam. The behavior of the plasma during the experiment is analyzed hundreds of times per second by complx calculations. However, it takes several minutes to perform each calculation which is only made available to physicists after conducting an experiment that occurs a few seconds only.

There is a decrease in the time to below 150 microseconds in the accurate prediction of the behavior of energetic particles through the new ML software. This allows calculations to be performed online during the experiment, according to the study.

The characteristics of plasma behavior not directly measured are estimated through the initial application of the model. This technique is a combination of predictions by ML and limited measurements of plasma conditions in real-time.