AI Enables Fast, Accurate Predictions for Better Control of Fusion Experiments

The software that drives self-driving cars and digital assistants is machine learning, a method employed in artificial intelligence (AI). It currently allows researchers to tackle the main challenges in harvesting on Earth the fusion energy that powers the stars and sun.

Recently, physicist Dan Boyer of the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) used this technology to design quick and precise forecasts for progressing control of experiments in the National Spherical Torus Experiment-Upgrade (NSTX-U) — the flagship fusion facility at PPPL that is under repair at present.

These AI predictions could help enhance the potential of NSTX-U researchers to improve the components of experiments that heat and shape the magnetically limited plasma that tends to fuel fusion experiments.

On improving the plasma’s heating and shaping, the researchers can study main aspects of the development of burning plasmas in a more effective manner — mostly self-heating fusion reactions that will be crucial for ITER, the international experiment that is under construction in France, and future fusion reactors.

Machine Learning Tactics

This is a step toward what we should do to optimize the actuators. Machine learning can turn historical data into a simple model that we can evaluate quickly enough to make decisions in the control room or even in real-time during an experiment.

Dan Boyer, Study Author and Physicist, U.S. Department of Energy, Princeton Plasma Physics Laboratory

Boyer is the author of a paper in Nuclear Fusion that explains machine learning tactics.

Light elements in the form of plasma have been integrated by fusion reactions — the hot, charged state of matter consisting of atomic nuclei and free electrons that makes up 99% of the visible universe — to produce enormous amounts of energy. The reproduction of fusion energy on Earth would make a virtually unlimited supply of safe and clean power to produce electricity.

Boyer together with his coauthor Jason Chadwick, an undergraduate student at Carnegie Mellon University and a Science Undergraduate Laboratory Internship (SULI) program participant at PPPL in summer 2020, tested machine learning predictions with the help of a decade’s data for NSTX, the forerunner of NSTX-U, as well as 10 weeks of operation of NSTX-U.

The shape of the two spherical tokamaks is more like cored apples compared to the doughnut-like shape of large and more extensively utilized traditional tokamaks, and they make economical magnetic fields that restrict the plasma.

The machine learning tests performed were able to accurately forecast the density of the electrons and distribution of pressure in fusion plasmas, where both were crucial but hard-to-predict parameters.

The electron pressure and density distribution within the plasma are key to understanding the behavior of fusion plasmas. We need models of these factors to predict the impact of changing heating and shaping on the performance and stability of experiments.

Dan Boyer, Study Author and Physicist, U.S. Department of Energy, Princeton Plasma Physics Laboratory

“While physics-based models for predicting electron pressure and density exist. They are not appropriate for real-time decision making. They take way too long to calculate and are not as accurate as we need them to be,” added Boyer.

Addresses Both Issues

Both the issues have been tackled by the machine learning model.

“It has learned to make predictions from thousands of observed profiles in the PPPL tokamaks and has made associations between combinations of inputs and outputs of actual data,” stated Boyer.

As soon as the model is trained, it takes less than one-thousandth of a second to assess. The speed of the consequent model could make it advantageous for several real-time applications.

However, the method has its own limitations.

“Since the model is trained on historically observed data, it cannot make predictions about new operating points with high accuracy,” added Boyer.

Boyer is planning to tackle this limitation by adding the results of physics-based model forecasts to the training data and developing methods of adjusting the model as new data becomes accessible.

Boyer has been awarded a highly competitive DOE five-year Early Career Research Program Award that will allow him and his collaborators to additionally expedite the optimization of the main components of fusion experiments.

This study was financially supported by the DOE Office of Science (FES). Assistance for Jason Chadwick comes from the DOE Science Undergraduate Laboratory Internships (SULI) program.

Journal Reference:

Boyer, M. D., et al. (2021) Prediction of electron density and pressure profile shapes on NSTX-U using neural networks. Nuclear Fusion. doi.org/10.1088/1741-4326/abe08b.