Experimental Plasma Discharge Dataset from the TCABR Tokamak

Description This dataset comprises 2,173 experimental plasma discharges from the TCABR (Tokamak à Chauffage Alfvén Brésilien) tokamak, operated by the Institute of Physics of the University of São Paulo (IFUSP). The database includes 1,684 non-disruptive and 489 disruptive discharges. All discharges are labeled for machine learning applications. The data features high-resolution (1 μs) magnetic diagnostics, including: Plasma current (Iₚ) – measured by Rogowski coils Loop voltage (Vₗₒₒₚ) – measured by toroidal loops Magnetic fluctuations – measured by an array of Mirnov coils distributed poloidally TCABR operates in a pure Ohmic L-mode regime with a circular cross-section, providing a highly reproducible environment (∼ 5% variation) free from confounding variables associated with auxiliary heating. This characteristic makes the dataset particularly valuable for studying MHD instabilities and disruption precursors, as well as for developing and validating machine learning-based disruption prediction algorithms. Classification Methodology Disruptive discharges were identified through the temporal evolution of the plasma current, focusing on its rapid terminal decay. An automated procedure detected candidate current-quench events by determining the times at which the plasma current crossed 80% and 20% of a local reference current, respectively, and calculating the corresponding decay interval, Δt80−20. When available, loop voltage, Mirnov coil signals, and Hα emission were used as complementary indicators. Finally, automated detections were manually reviewed to distinguish disruptive events from programmed current ramp-downs, incomplete records, and other operational conditions exhibiting similar signal characteristics. Technical Validation The data underwent rigorous validation, including: Calibration traceable to INMETRO (Brazilian National Institute of Metrology, Quality and Technology) Cross-diagnostic consistency checks Monte Carlo sensitivity analysis with ±20% uniform perturbation of classification thresholds, demonstrating label robustness for 99.9% of discharge Data Format Data are available in NetCDF format (tcabr_data.nc), organized in a hierarchical structure by shot ID. We recommend using the xarray Python library for data manipulation and analysis. Each shot contains multiple diagnostics, which are listed in the table below: Diagnostic Name Diagnostic Alias Mirnov coils 1--24 BbMirnovN01-N24 Flux coil BobFlux Toroidal magnetic field CpToroidal Electrode current EletrCurrent Electrode voltage EletrVoltage Gas injection valve GasPuffing02 Hard X-Ray HardXRay Hα HAlfaRef Plasma current IPlasma Vertical field current 01 IVert01 Vertical field current 02 IVert02 Loop voltage VLoop Sinusoidal coils voltage VSin A Python script (tcabr_tools.py) is also provided, containing the custom implementations used for the disruption labeling algorithm, Monte Carlo sensitivity analysis, and data analysis. In addition, a machine learning notebook (tcabr_ml.ipynb) is included, demonstrating the training and evaluation of k-Nearest Neighbors (kNN), Random Forest, and XGBoost classifiers using the provided dataset. TCABR Mirnov coil geometry and NetCDF channel mapping A CSV file (tcabr_mirnov_coils.csv) is also provided, containing geometric information about the TCABR Mirnov coils. The installed array comprises 22 poloidal coils B0–B21 at phi=0 degrees and two additional coils B3a (phi=112 degrees) and B3b (phi=220 degrees). Channel mapping confirmed by the TCABR team: B0–B21 correspond, in order, to BbMirnovN01–BbMirnovN22; B3a to BbMirnovN23; B3b to BbMirnovN24. R_m and Z_m are in metres; theta_deg, theta_star_deg and phi_deg in degrees; each coil has 117 turns and area_m2=6.615e-5. Coordinates and angles are transcribed from the confirmed installed-geometry table; conversion factors from MATLAB geometric documentation. The physical naming-to-channel correspondence is based on the TCABR team confirmation, not independently checked against MDSplus wiring configurations. Differences Between Versions The tcabr_manifest.csv file provides a comparison of the plasma discharges available in dataset versions 1.0.0 and 1.1.0 for traceability purposes. To ensure consistency and access to the most up-to-date information, users are strongly encouraged to use only the latest version of the dataset (version 1.1.0). Potential Applications Training and validation of machine learning models for disruption prediction Studies of MHD instabilities (e.g., m/n = 2/1 modes) Transport and confinement analysis in Ohmic plasma Benchmarking for numerical simulation codes Development of real-time disruption mitigation systems Changelog v1.1.0 — 2026-10-31 — Included a Machine Learning notebook for discharge classification, new Python scripts, Mirnov coil geometry information, a manifest tracking shot changes between versions 1.0.0 and 1.1.0, and updated NetCDF file. v1.0.0 — 2026-08-09 — Initial Release