Identification of Alfvén eigenmodes using recurrent neural networks, a labelled database and CO₂ interferometer data on DIII-D

Artificial Intelligence is developed to automatically detect Alfvén eigenmodes (AE) and these models achieve high performance (True Positive Rate = 90% and False Positive Rate = 14%). Using labels created from a curated database [Heidbrink, et al., NF ‘20], Machine Learning-based systems are trained using single chord and crosspower spectrograms to predict the presence 5 AEs (EAE, TAE, RSAE, BAE and LFM). Since resonant fast ions can drive AEs unstable and degrade the plasma performance or energy confinement, this work demonstrates the potential of applying Machine Learning methods to detect and control AEs. In this work, the advantages of using the CO₂ interferometer to detect AEs, and the results from a comparison between inputs (single chord and crosspower spectrograms) and another comparison between two different models (Reservoir Computing Network and Long Short-Term Memory Network) are presented. The highest performance is achieved by the Reservoir Computing Network trained with single chord spectrograms. Also, AE detection using any chord is feasible (the vertical chord passing near center is best). These models can be implemented into control algorithms that drive actuators for the mitigation of unwanted AE impacts. Supported by the U.S. Department of Energy under DE-FC02-04ER54698, DE-SC0021275, DE-SC0020337, DE-SC0014664, Army Research Office (ARO W911NF-19-1-0045), National Science Foundation under 1633631 and Ghent University Special Research Award No. BOF19/PDO/134.