Real-Time Detection of Confinement Regimes in Fusion Plasmas via Deep Learning and Edge Turbulence Measurements

Real-time detection of the plasma confinement regime can offer new capabilities for achieving and maintaining enhanced confinement states with advanced plasma control. To demonstrate this, we employed the 2D beam emission spectroscopy (BES) diagnostic system, capturing localized density fluctuations of long-wavelength turbulent modes in the edge region at a 1 MHz sampling rate. BES data from 330 discharges—spanning L-mode, H-mode, quiescent H (QH)-mode, and wide-pedestal QH-mode—was collected from the DIII-D tokamak to develop a robust database for training a deep-learning classification model. A shallow 3D convolutional neural network was employed for the multivariate time-series classification task, which generates a probability distribution over the different confinement regimes. Our model achieved an average accuracy of 0.95 and an average F₁ score of 0.94 on 44 unseen test discharges, using only ~1 millisecond snippets of BES data at a time. Furthermore, we explored what features of the turbulence are most crucial to good model performance by varying the model input dimensionality. This study demonstrates the feasibility of real-time analysis of high-bandwidth fluctuation data in future devices like ITER, where advanced and reliable plasma control is imperative.