May 20 2019
Machine learning (ML) is a kind of artificial intelligence that understands language, recognizes faces, and navigates self-driving cars. It could be helpful in bringing the clean fusion energy that lights the sun and stars to Earth.
Fast-camera photo of a plasma produced by the first NSTX-U operations campaign. (Image credit: Photo by NSTX-U experiment)
Scientists from the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) have been using ML to develop a model for rapid control of plasma that fuels fusion reactions. Plasma is the state of matter formed of free electrons and ions, or atomic nuclei.
The sun and a majority of the stars are colossal balls of plasma that undergo constant fusion reactions. However, on Earth, researchers must heat and control the plasma to make the particles to fuse and liberate their energy. The study by PPPL demonstrates that ML can enable such control.
Neural Networks
Scientists headed by Dan Boyer, a PPPL physicist, have trained neural networks—the crux of ML software—on data synthesized in the first operational campaign of the National Spherical Torus Experiment-Upgrade (NSTX-U), the flagship fusion facility, or tokamak, at PPPL. The trained model can accurately reproduce the predictions of the behavior of the energetic particles synthesized by powerful neutral beam injection (NBI) used to fuel NSTX-U plasmas and heat them to million-degree, fusion-relevant temperatures.
In general, these predictions are performed by a complex computer code known as NUBEAM, incorporating information related to the effect of the beam on the plasma. It is necessary for such complex calculations to be made hundreds of times within a second to study the behavior of the plasma at the time of an experiment. However, several minutes can be needed for each calculation to be performed, rendering the results available to physicists only following an experiment that usually lasts a few seconds.
The new ML software minimizes the time required for the accurate prediction of the behavior of energetic particles to less than 150 μs, thereby allowing the calculations to be performed online at the time of the experiment.
Preliminary application of the model showed a method for predicting the properties of the plasma behavior not measured directly. This method integrates ML predictions with the restricted measurements of plasma conditions available in real time. The integrated outcomes will assist the real-time plasma control system in making more informed decisions related to the way to adjust beam injection to optimize performance and preserve the stability of the plasma—a vital quality for fusion reactions.
Rapid Evaluations
The rapid assessment will also enable operators to make more knowledgeable adjustments between experiments that are performed every 15–20 minutes at the time of operations.
Accelerated modeling capabilities could show operators how to adjust NBI settings to improve the next experiment.
Dan Boyer, Physicist, PPPL
Boyer is also the lead author of a paper published in Nuclear Fusion that describes the new model.
Boyer collaborated with Stan Kaye, another PPPL physicist, created a database of NUBEAM calculations for a variety of plasma conditions analogous to those realized in experiments at the time of the preliminary NSTX-U run. The database was used by the scientists to train a neural network to estimate the impacts of neutral beams on the plasma, for example, heating and profiles of the current. Subsequently, Keith Erickson, a software engineer, executed the software for assessing the model on computers used to actively control the experiment to test the calculation time.
The new study will involve the creation of neural network models customized to the planned conditions of forthcoming NSTX-U campaigns and other fusion facilities. Furthermore, researchers intend to widen the existing modeling strategy to allow more rapid predictions of other fusion plasma phenomena. The DOE Office of Science supported this study.