Detection of Alfvén Eigenmodes on COMPASS with Generative Neural Networks

Abstract

Chirping Alfvén eigenmodes were observed at the COMPASS tokamak. They are believed to be driven by runaway electrons (REs), and as such, they provide a unique opportunity to study the physics of nonlinear interaction between REs and electromagnetic instabilities, including important topics of RE mitigation and losses. On COMPASS, they can be detected from spectrograms of certain magnetic probes. So far, their detection has required much manual effort since they occur rarely. We strive to automate this process using machine learning techniques based on generative neural networks. We present two different models that are trained using a smaller, manually labeled database and a larger unlabeled database from COMPASS experiments. In a number of experiments, we demonstrate that our approach is a viable option for automated detection of rare instabilities in tokamak plasma.

Promoting global stability in data-driven models of quadratic nonlinear dynamics

“…The proper orthogonal decomposition (POD) [10,66] is frequently used to obtain the basis, since the modes ϕ i (x) are orthogonal. Many other modal expansions and bases have been introduced for reduced-order fluid [8,9] and plasma models [67][68][69][70], including balanced POD [71,72], spectral POD [20], dynamic mode decomposition (DMD) [11,12,73], the Koopman decomposition [13,74,75], resolvent analysis [76,77], and autoencoders [32,78,79]. The governing equations are then Galerkin projected onto the basis {ϕ i (x)} by substituting Eq.…”

Comparison of Anomaly Detectors: Context Matters

“…A recurring idea [25,94,74,81] is to combine autoencoders with a secondary model acting on the latent space defined by the encoder. The rationale behind it is the encoder preserves semantic information of the sample and removes noise (e.g.…”

“…In [74], the model optimizes the projection of data (by virtue of NNs) to a new space, where they can be easily enclosed in a sphere of minimum radius. Approach presented in [81,94] explicitly splits the creation of the detector into two parts. It first trains a VAE (and its variants) and then it fixes the encoder.…”

“…It first trains a VAE (and its variants) and then it fixes the encoder. The anomaly score is calculated by a kNN (in [81]) or by OC-SVM (in [94] detectors in the latent space, obtained by projecting the sample by the fixed encoder. The two-stage models can be also viewed as a kNN with a trained metric or OC-SVM with a trained kernel.…”

Alfvén eigenmode classification based on ECE diagnostics at DIII-D using deep recurrent neural networks

“…There has already been remarkable success in machine learning for disruption identification and realtime control in tokamaks [5,6,[27][28][29][30], including highperformance models that are not limited to a specific device [26]. There has also been recent deep learning work for magnetohydrodynamic (MHD) and Alfvén eigenmode activity, which utilized manually-labeled spectrogram data from the TJ-II stellarator [31] and COM-PASS tokamak [32] for automated identification of these modes in diagnostic data from a single magnetic probe. The former paper focuses on a binary classification of the spectrogram pixels, indicating whether each pixel corresponds to Alfvénic MHD activity or not.…”