Deep neural network based real-time ELM prediction and reconstruction of turbulent flow based on the DIII-D BES measurement

Since ELM crashes cause losses and fast relaxation of the pedestal energy and profiles, it is crucial to predict them beforehand to achieve high-performance and safe tokamak operations. With 2D turbulence and eddy flow measurements of electron density given by the DIII-D BES system, we develop a neural network which is able to predict not only ELMs several msecs before their onsets but the first ELMs after the L-H transition under about 0.8 true positive rate. Since BES configuration with a time window containing the spatiotemporally fluctuating density is used as the network inputs, it is also necessary to know what information the network obtains from its input. Thus, we train another network to prove that the turbulent velocity fields are one of the information being able to be captured by the network. As a preliminary result, we use the time-delayed cross-correlation to generate the ground truth of radial velocity profiles as a training dataset. Therefore, we expect that our network is possible to extract the velocity information from the BES signals and provide various velocity-related characteristics of plasmas in real time.