Tearing Mode Database v0

Tearing Mode Database V0

This is a PRELIMINARY release of the tearing mode database (TMDB), containing 118 shots of data from the Alcator C-Mod tokamak during the 2010-2016 campaigns. The labeling procedure is highly automated and has the potential to classify any periodic rotating magnetic perturbation as an NTM. Its contents have been curated to the best ability of a single person, and has yet to undergo rigorous peer and code review. As such there may be inaccuracies and the organization structure is subject to change. The purpose of this release is to demonstrate the capabilities of the analysis methods, get feedback, and iterate prior to a more comprehensive dataset in the future. All questions/comments should be forwarded to [e-mail] and [e-mail].

Data Sources

All data from Alcator C-Mod was obtained with the Disruption-Py library [1][2].

The 1 kHz EFIT computation was performed with disruption-efit [3].

The C-Mod vessel model was created by Rian Chandra.

Dataset Contents

All signals are mapped to a uniform timebase at 1 kHz with interpolation that maintains causality.

Since the NTM labeling requires the flux surface shapes from an equilibrium reconstruction, the NTM signals are NaN whenever the reconstruction is missing for a given timestep.

The EQDSK names follow the conventions as given by the FreeQDSK library.

For a full list of signals and their descriptions, see the attached `dataset_description.md`. 

NTM Labeling Methods

Our method of labeling NTMs candidates centers around spectral analysis. The measured cross-phases between multiple high-frequency (2.5 MHz) Mirnov sensors are compared against the cross-phases of synthetic sensors to determine which mode structure produces the best fit. The `ntm_sxx` signal is determined by the peak spectral power density from a Mirnov sensor at a given frequency. This is then corrected by removing the contribution of induced eddy currents in the vacuum vessel to produce `ntm_amp`.

The cross-phase from the measured data is determined from the short-time fast Fourier transform of each sensor. To match the 1 kHz timebase without interpolation, the frequency bins are also 1 kHz wide. The Fourier transform results for a single probe are stored in `mirnov_fft_real` and `mirnov_fft_imag`.

To determine the general location of an NTM candidate, we use the WavyStar library [4]. This analyzes the spectrogram from a single Mirnov probe and uses image segmentation to isolate a rotating perturbation's position in time and frequency space.

For each timestep an NTM may exist, we trace filaments on the rational surface corresponding to a particular mode structure and simulate the signal from the Mirnov sensors. Synthetic Mirnov data is calculated with the ThinCurr library [5] (part of the OpenFUSIONToolkit). The phase differences between measured signals and synthetic signals are compared across 10 candidate modes. The chosen mode is the one with the minimum $\chi^2$ value as determined by $\chi^2 = \frac{\sum_{i=1}^{N} (\phi_\text{meas} - \phi_\text{synth})^2}{N}$, where $N$ is the number of active Mirnov sensors. This process uses a simple vessel model for performance reasons, as we have determined that the vessel response does not meaningfully impact the synthetic phase at NTM frequencies under 100 kHz.

Once a particular NTM structure has been identified at a point in frequency space, we run ThinCurr again, this time using a detailed vessel model to correct for the eddy currents induced in the vacuum vessel.

After this automated labeling procedure, the results are curated by a human to ensure the labeled NTM candidates qualitatively exhibit NTM-like behavior.