Multi-Diagnostic Classification of Alfvén Eigenmodes using Multimodal Machine Learning

Real-time Machine Learning based control on tokamaks requires efficient data processing and featurization of high-frequency diagnostics. By using multiple high-frequency diagnostics, we can create a multi-modal representation that yields a better classification of Alfvén Eigenmode (AE) activity. Previous work has used a database of 1000 labeled DIII-D shots to look at individual Electron Cyclotron Emission (ECE) channels and classify four main types of AE activity: Low-frequency modes (LFMs), Beta-induced Alfvén eigenmodes (BAE), Reversed-Shear Alfvén eigenmodes (RSAE), and Toroidal Alfvén eigenmodes (TAE).

We build upon this work and show improved classification as well as prediction of this AE activity by using a multi-diagnostic model. In addition to ECE, we use cross-power CO2 Interferometry, Beam Emission Spectroscopy (BES), and other magnetic diagnostics in our model. Additionally, these diagnostics have a significant amount of noise present and require pre-processing to reduce noise in the spectrograms. We utilize a source-invariant denoising autoencoder to clean the spectrograms of all the diagnostics.