Building Fusion Control with Machine Learning
Applied Physics doctoral student Yumou "William" Wei published his latest research on controlling a fusion plasma using machine learning. The paper was invited as part of the Special Issue on the 26th Workshop on MHD Stability Control, [Y. Wei, et al 2023 Plasma Phys. Control. Fusion 65 074002; https://iopscience.iop.org/article/10.1088/1361-6587/acd581], and was co-authored by Drs. Jeff Levesque and Chris Hansen and by Profs. Mike Mauel and Gerald Navratil.
William used dual high-speed video cameras to detect light fluctuations from the plasma as instabilities caused the hot plasma to graze the cold wall surrounding the plasma. Then using a state-of-the-art graphical processing unit (called a "GPU"), he implemented a new algorithm to track the amplitude and phase of the instabilities trained by artificial intelligence, or "deep learning.” For the first time, William showed how a neural network could successfully predict key parameters using solely optical measurements from one or more cameras. The new algorithm outperformed other, more conventional, algorithms, and William explored the impact of different input data streams on the accuracy of the model's predictions.
This work was carried our on the High Beta Tokamak - Extended Pulse (HBT-EP) device. Additional design and optimization is underway for deploying neural network models on field-programmable gate arrays (FPGA) which will further speed-up the rate of mode identification and satisfy the requirements for real-time mode feedback control on HBT-EP and other fusion energy devices.
This work was supported by U.S. Department of Energy, Office of Science, Office of Fusion Energy Science, Grant No. DE- FG02-86ER53222.