Low latency optical-based mode tracking with machine learning deployed on FPGAs on a tokamak

Area of Science:

• Plasma physics and fusion energy research.
• Advanced diagnostics and real-time control systems.
• Machine learning applications in scientific instrumentation.

Background:

• Active feedback control is crucial for stabilizing plasma instabilities in magnetic confinement fusion devices.
• High-speed optical cameras offer non-invasive plasma monitoring capabilities.
• Real-time processing of diagnostic data is essential for effective control.

Purpose of the Study:

• To develop and demonstrate an in situ FPGA-based system for processing high-speed camera data.
• To track magnetohydrodynamic (MHD) mode evolution and generate control signals in real time.
• To leverage deep learning for enhanced accuracy in plasma mode prediction.

Main Methods:

• Utilized high-speed cameras operating at over 100 kfps.
• Implemented a convolutional neural network (CNN) model on in situ field-programmable gate array (FPGA) hardware.
• Integrated the CNN model within the camera's FPGA readout for low-latency processing.

Main Results:

• Achieved real-time tracking of magnetohydrodynamic (MHD) mode amplitude and phase.
• CNN model demonstrated superior accuracy in predicting MHD modes compared to non-deep-learning methods.
• System achieved a total trigger-to-output latency of 17.6 μs and a throughput of 120 kfps.

Conclusions:

• The FPGA-based high-speed camera system enables real-time, machine-learning-driven tokamak diagnostics and control.
• Demonstrated the feasibility of integrating advanced AI models into fusion diagnostic hardware.
• The system holds potential for broader applications in scientific domains requiring high-speed data processing and analysis.