Spatially-localized Alfvén eigenmode classification using convolutional neural networks

We use an expert-labeled database of DIII-D discharges to classify five types of Alfvén eigenmodes (AEs) with convolutional neural networks, opening up the possibility of deep-learning-enhanced real-time control of this important class of plasma dynamics. Each DIII-D discharge in the database consists of forty radially-localized electron cyclotron emission (ECE) measurements, sampled at 500 kHz for the first 2 seconds of the discharge. The model attempts to predict when each AE type occurs in a validation dataset, and discriminates between the five types of AE activity. This strategy performs strongly at spatio-temporally localized prediction and classification of Alfven eigenmodes (approximate average true positive rates of 80% and false positive rates of 2%), indicating future promise for more sophisticated spatio-temporal models and incorporation into future real-time control strategies.