Leveraging the MHz-scale 2D BES diagnostic in real time to predict the ELM onset based on deep learning acceleration

Understanding, predicting and controlling of high frequency plasma dynamics such as turbulent flow dynamics and plasma fluctuations in real time can provide information for maintaining stability in tokamak plasmas. This can be achieved by incorporating high bandwidth MHz-scale fluctuation diagnostics into the real-time analysis and tokamak control, thereby enhancing our knowledge of the plasma control system for future fusion reactors. Recently, a deep-learning based analysis for the high bandwidth 2D BES system was developed, which enables us to forecast the onsets of the ELMs in the H-modes on the order of 10-100 ms beforehand by examining 100 μs data window from the data stream. Here, this newly suggested method is extended into various operation regimes such as the QH-mode to effectively predict so-called breakthrough ELMs in the ELM-free and suppressed regimes; moreover, the method is tested with the real-time BES system in DIII-D. The convolutional layers are optimized and quantized as simple as possible for the FPGA acceleration, allowing for hundreds of μs scale latency, without losing the network capability. This model allows us to interpret itself through the activation features, showing that learning the fluctuations along the poloidal direction is essential to identify ELM relevant and irrelevant information.