FPGA-based microsecond-latency MHD mode tracking using high-speed cameras and deep learning on HBT-EP

A deep-learning-based MHD mode tracking algorithm using high-speed imaging cameras has been developed for real-time feedback control application on the High Beta Tokamak – Extended Pulse (HBT-EP) device. Our algorithm [1] uses the convolutional neural network (CNN) to process video frames taken during plasma shots by one or multiple cameras and predict the n=1 mode amplitude and phase over time. The model is able to accurately track the n=1 modes consistently over the testing shots and demonstrated significant improvement over the previous SVD-based method [2].

For real-time application we utilize the hls4ml (High Level Synthesis for Machine Learning) [3] framework to optimize the deep learning models for deployment onto Xilinx FPGA devices. Using hls4ml, our model is implemented directly on the Euresys Coaxlink Octo framegrabber board in the existing camera diagnostics system and achieves an input-to-output latency below 17 μs, on par with the current GPU-based control system using the magnetic sensors. The proposed controller will later be integrated into the feedback control system on HBT-EP to perform real-time mode control.

[1] Wei, Y. et al (2023) PPCF 65 074002

[2] Angelini, S. et al (2015) PPCF 57 045008

[3] Fahim, F. et al (2021), arXiv:2103.05579