UKAEA-IBM-STFC/tokamark-v1

TokaMark

TokaMark is a large, open benchmark dataset for evaluating AI models on real experimental diagnostics from the Mega Ampere Spherical Tokamak (MAST). It packages 39 harmonized signals from 11,573 MAST shots into a standardized, task-ready format spanning 14 downstream tasks across plasma equilibrium reconstruction, short-term magnetics dynamics, kinetic profile evolution, and long-range forecasting of MHD activity and disruptions.

This repository hosts the data release accompanying the paper:

Cécile Rousseau, Samuel Jackson, Rodrigo H. Ordonez-Hurtado, Nicola C. Amorisco, Tobia Boschi, George K. Holt, Andrea Loreti, Eszter Székely, Alexander Whittle, Adriano Agnello, Stanislas Pamela, Alessandra Pascale, Robert Akers, Juan Bernabe Moreno, Sue Thorne, Mykhaylo Zayats. "TokaMark: A Comprehensive Benchmark for MAST Tokamak Plasma Models." arXiv:2602.10132, 2026.

Code, task definitions, baseline models, and evaluation tooling are released at UKAEA-IBM-STFC-Fusion-FMs/tokamark_baseline.

Dataset summary

TokaMark is built on top of FAIR-MAST, the only openly available dataset of real tokamak diagnostics, curated from the last five experimental campaigns of MAST (operated by UKAEA/EURATOM in Culham, UK, 1999–2013). From FAIR-MAST's full archive, TokaMark selects and freezes a stable subset of 39 signals across 11,573 shots, harmonizes metadata and units, and organizes them into standardized input/output windows for 14 benchmark tasks.

Signals are grouped along four axes:

Category	Example signals	Origin	Frequency	Modality
Magnetics	Flux loops, pickup coils, saddle coils	Diagnostic	5 kHz; 50 kHz	Profile
Kinetics	Thomson scattering, interferometer	Diagnostic	0.2 kHz; 4 kHz	Profile; time series
Radiatives	D-alpha, soft X-ray	Diagnostic	50 kHz	Profile
Fast magnetics	Mirnov coils	Diagnostic	500 kHz	Profile
Currents	Poloidal field coil currents, solenoid/plasma current	Diagnostic	4 kHz	Profile; time series
Voltages	Poloidal field coil voltages	Actuator	4 kHz	Profile
References	Reference plasma current/density	Actuator	4 kHz	Time series
Fueling	NBI power, gas puffing	Actuator	4 kHz	Time series
Equilibrium	Shape parameters, J_tor metrics, flux map	Derived (EFIT)	0.2 kHz	Time series; profile; video

A full signal-level listing is provided in the paper's Appendix A.

Tasks

TokaMark defines 14 tasks organized into 4 groups:

Group	Tasks	Objective	Input window	Output window
1 — Instantaneous Equilibrium Reconstruction	1-1, 1-2, 1-3	Reconstruct plasma equilibrium (shape parameters, boundary, flux map) from magnetics and coil currents	5 ms	5 ms
2 — Short-Term Magnetics Dynamics	2-1, 2-2, 2-3	Forecast currents, equilibrium metrics, boundary, and flux map in response to actuator commands	5 ms	25 ms
3 — Kinetic Profile Dynamics	3-1, 3-2, 3-3	Forecast/reconstruct Thomson scattering, radiatives, and equilibrium quantities from partial, multi-rate diagnostics	5 ms	50 ms
4 — Long-Range Forecasting of MHD Activity	4-1 – 4-5	Long-horizon forecasting of soft X-ray, shape, plasma current, and Mirnov (spectral) signals ahead of disruptions/locked modes	100 ms	—

Each task is a Reconstruction, Autoregressive (AR) Forecasting, or Reconstructive (RC) Forecasting problem, and some (Group 3 Task 3-3, all of Group 4) additionally require Non-Markovian (long) input histories. Full task definitions, input/output signal sets, and the hierarchical (samples → windows → signals → tasks → shots) NRMSE evaluation protocol are described in Sections 2–3 of the paper.

Data splits

The benchmark provides two splitting strategies, both performed at the shot level to avoid leakage:

• Random split — 80/10/10 train/val/test.
• Temporal split — the two most recent experimental campaigns are held out as the test set, to assess generalization across campaigns/time.

Per-task shot counts for both strategies are listed in Appendix B of the paper (roughly 5,500–9,000 shots per task, depending on missing-signal filtering).

Dataset structure

The raw data is distributed as a set of batch_XXXXX.tar archives (~565 GB total). Each tar batch contains multiple zipped Zarr stores (*.zarr.zip), one per shot, following FAIR-MAST's per-shot Zarr layout, where each signal is identified as <group_name>-<signal_name> (e.g. equilibrium-psi, magnetics-flux_loop_flux, thomson_scattering-t_e).

This packing (tar of zipped Zarr stores) is only used to make the ~11.5k individual shot stores efficient to host and download on the Hub — you need to unpack it after downloading before the data is usable.

Downloading and unpacking

1. Download all batch_XXXXX.tar files (and unpack_all.py) from this repository, e.g. with the huggingface_hub CLI:

hf download UKAEA-IBM-STFC/tokamark-v1 --repo-type dataset --local-dir ./tokamark-v1

2. Run the included unpack script to restore the individual .zarr stores. It reverses the packing pipeline in two steps: extracting each batch_XXXXX.tar to recover the per-shot .zarr.zip files, then unzipping each of those into a real .zarr directory.

python unpack_all.py ./tokamark-v1 ./restored_zarrs --workers 8

• batches_dir — folder containing the downloaded batch_XXXXX.tar files.
• output_dir — destination folder for the restored .zarr stores.
• --workers — number of parallel processes (defaults to CPU count).
• --keep-intermediate-zips — optionally keep the extracted per-shot .zarr.zip files instead of deleting them once unzipped.

The script prints progress as batches are extracted and stores are unzipped, and reports any failures individually rather than aborting the whole run.

After unpacking, restored_zarrs/ will contain one .zarr directory per shot, ready to be loaded (e.g. with zarr/xarray) or fed into the task-specific data loaders from the tokamark_baseline toolkit.

Note: unpacking requires roughly as much free disk space again as the downloaded archive size, since both the tarred and unpacked copies exist during the process (more if you keep intermediate zips).

Baselines

The paper provides four baselines evaluated on all 14 tasks: Mean and Persistence (naive statistical models), and a multi-branch convolutional encoder–decoder with two variants, CNN and CNN+LSTM. Results (Task/Group-level NRMSE, lower is better) are reported for both random and temporal splitting in Table 3 of the paper, showing that deep baselines substantially outperform naive ones under random splitting, but that this advantage degrades — and can reverse — for profile-dynamics and MHD-activity tasks under the more realistic temporal split.

Intended uses

TokaMark is intended for benchmarking AI/ML models — from task-specific architectures to foundation models pretrained across tasks — on real, heterogeneous, and imperfect fusion diagnostic data. It targets representation learning, multi-rate/multi-modal fusion, robustness to missing data, and generalization across operating regimes and campaigns.

Citation

@article{rousseau2026tokamark,
  title={TokaMark: A Comprehensive Benchmark for MAST Tokamak Plasma Models},
  author={Rousseau, C{\'e}cile and Jackson, Samuel and Ordonez-Hurtado, Rodrigo H. and Amorisco, Nicola C. and Boschi, Tobia and Holt, George K and Loreti, Andrea and Sz{\'e}kely, Eszter and Whittle, Alexander and Agnello, Adriano and Pamela, Stanislas and Pascale, Alessandra and Akers, Robert and Bernabe Moreno, Juan and Thorne, Sue and Zayats, Mykhaylo},
  journal={arXiv preprint arXiv:2602.10132},
  year={2026}
}

Please also consider citing the underlying FAIR-MAST data releases:

• Jackson, S. et al. (2024). FAIR-MAST: A fusion device data management system. SoftwareX 27, 101869.
• Jackson, S. et al. (2025). An Open Data Service for Supporting Research in Machine Learning on Tokamak Data. IEEE Transactions on Plasma Science.

And consider citing the relevant MAST result publications:

• Sykes, A., et al. (2001). First results from MAST. Nuclear Fusion.
• Morris, A. W. et al. (2011). MAST: results and upgrade activities. IEEE/NPSS 24th Symposium on Fusion Engineering.

Acknowledgements / provenance

Data derived from MAST diagnostics collected by UKAEA and EURATOM, made openly available via FAIR-MAST. This benchmark packaging is a collaboration between IBM Research Europe, UK Atomic Energy Authority (UKAEA), and the STFC Hartree Centre.

License

Released under CC BY 4.0, consistent with the accompanying paper. Please verify license terms for any downstream redistribution given the underlying FAIR-MAST data provenance.