Data

The dataset

The first open-access, harmonized multi-machine benchmark for fusion equilibrium — curated for Thomson-diagnostic availability, feature completeness, and EFIT-reconstruction quality.

The challenge releases 7,915 DIII-D shots (7,041 train + 874 public test) and 1,206 MAST shots (zero-shot transfer set) — 9,121 shots, 98 GB total, each filtered so that the inputs and EFIT ground truth a model needs are actually present and reconstruction-quality. The full corpus is hosted on Hugging Face; the datasets library supports streaming.

Every shot is one row of a Parquet file. Each time series and profile is a nested array inside that single row — so df['efit_psirz'].iloc[0] is a list of 2D flux grids, not a column of scalars. Always take .iloc[0], then index into the nested array.

Six sample shots (3 DIII-D, 3 MAST) ship in the starter kit so you can explore the schema immediately, without downloading the full corpus.

Data pipeline: raw multi-rate diagnostics harmonized into one-row-per-shot Parquet records.
Multi-rate diagnostics harmonized into one row per shot.
Dataset (Hugging Face) Quickstart →
Read this before you model

Every plasma in the dataset is diverted

A tokamak plasma's boundary is set one of two ways: by a magnetic X-point X-point The point where magnetic surfaces cross in a figure-of-eight. Its presence means the plasma is diverted (exhaust steered to a target); its absence means limited (plasma resting against a wall). (diverted), or by physical contact with a material surface (limited). This release contains diverted frames only — limited frames were removed from efit_times, and from every column that lives on it, on both machines and all splits.

That is what makes the problem well-posed. On a limited plasma the last closed flux surface is determined by the wall, and wall geometry is not contained in ψ(R,Z) Poloidal flux ψ(R,Z) The 2D field you are asked to predict. Its contour lines are the magnetic surfaces that hold the plasma in place — so a map of ψ is effectively a picture of the plasma’s shape. — so a boundary extracted from the flux map alone systematically overshoots. Scoring a boundary you cannot infer is not a useful test, so those frames are gone rather than quietly down-weighted. One consequence for scoring: no shape scalar is skipped for being on a limited frame any more, because there are none.

efit_times is no longer contiguous. Dropping frames leaves gaps in the target time base, so do not assume a uniform timestep. The full-rate inputs (magnetics_*, thomson_*) are not subset — you still get the complete discharge and resample it onto efit_times exactly as before.
Two machines

What differs between DIII-D and MAST

The differences below are load-bearing for any code touching the data.

DIII-D (conventional) MAST (spherical)
Type Conventional tokamak (D-shaped) Spherical tokamak (low aspect ratio)
Facility DIII-D · General Atomics, San Diego MAST · UKAEA Culham, UK
Shots (released set) 7,915 1,206
Training data 7,041 shots None — zero-shot by design
Target flux grid 65 × 65 65 × 65
Shaping coils 18 F-coils (F1A–F9B) + ECOILA, bcoil 10 P-coils (P2L–P6U) + sol, tf, efps
Magnetics time base Shared magnetics_time, length varies by shot (480k @ 20 kHz, or ~49k @ 2 kHz); Ip on its own ADC Shared magnetics_time (~30k samples)
Typical flux value ≈ −0.25 V·s/rad ≈ +0.05 V·s/rad
What’s inside a shot

Signals & targets

Target — EFIT ψ(R,Z)

Predict this
efit_psirz: a sequence of poloidal flux maps Poloidal flux ψ(R,Z) The 2D field you are asked to predict. Its contour lines are the magnetic surfaces that hold the plasma in place — so a map of ψ is effectively a picture of the plasma’s shape. (one per efit_times slice) — the ground truth your model reconstructs, kept exactly as EFIT EFIT Equilibrium Fitting — the standard code that reconstructs a tokamak’s equilibrium from its sensors. Its output is the ground truth you are scored against; the whole point of the challenge is to reach it without the magnetic sensors EFIT relies on. produced it. The EFIT scalar labels are additional target quantities, withheld on the test splits; you submit only q95 and betaN — the geometric ones are derived from your flux map at scoring.

Magnetics — coil currents

Input
F-coils / P-coils Coil currents The currents in the machine’s electromagnets — the "control panel" of the tokamak. The other input family. DIII-D has 18 F-coils plus ECOILA and bcoil; MAST has 10 P-coils plus solenoid, TF and EFPS. , ohmic solenoid, toroidal-field coil, and plasma current ( Ip Ip (plasma current) The total electric current flowing through the plasma itself — of order 10⁶ A, roughly 100× the coil currents. Given as an input. ). Sampled at tens of kHz on per-signal (DIII-D) or shared (MAST) time bases. Shaping-coil currents are in ampere-turns (kA·turn) on both machines, so they are directly comparable across machines. ( dsep dsep EFIT-derived context shipped in train and withheld on test — not an input, and not scored. On DIII-D it is EFIT’s a-file DSEP, a separatrix-to-limiter clearance whose sign encodes the configuration (> 0 diverted): it is the criterion that defined this diverted-only corpus. On MAST the same column is a different physical quantity — esm/dr_sep_out, the upper/lower divertor balance — and its sign is not a diverted/limited indicator. Because the two machines’ columns do not mean the same thing, dsep was dropped from the scored metric. is EFIT-derived and is context, not an input — and it is not scored.)

Thomson scattering

Input
Core and edge Thomson scattering A laser diagnostic: fire a laser through the plasma and read the scattered light to get electron temperature (Te) and density (ne). One of the two input families you are given — and, crucially, not a magnetic sensor. electron temperature Te (eV) and density ne (m⁻³) profiles, on their own time bases. DIII-D's core laser is vertical and MAST's is a horizontal midplane chord.
Machine geometry ships with every row. A coil current only means something once you know where the coil is, so each shot carries the conductor rectangles (coil_R, coil_Z, coil_width, coil_height, plus skew angles for the six DIII-D parallelogram F-coils) and the Thomson chord positions (thomson_chord_R, thomson_chord_Z). Join geometry to currents on coil_input_column. These are inputs — nothing here is withheld on the test splits.
Read this first

Modeling rules the data treats as load-bearing

Align time bases — but only the inputs

Magnetics (~49k samples on DIII-D, ~30k on MAST), the EFIT targets (a few hundred frames), and Thomson all live on different clocks. Resample inputs onto EFIT times; never interpolate the targets.

Split by shot, never by timestep

Timesteps within a shot are highly correlated. Splitting by timestep leaks the test set; always split at the shot level.

Normalize your inputs

Signal scales span ~10⁴ A coil currents to ~10⁶ A plasma current. Normalize before training.

Compress the target

The 4,225-pixel flux map compresses to ~20–50 PCA coefficients (component 1 alone ≈ 92% of variance) — the recommended baseline for strong generalization.
One target grid on both machines. efit_psirz is a 65×65 poloidal flux map on a physical (R, Z) grid in metres (efit_grid_R, efit_grid_Z) — the same shape for DIII-D and MAST, so a model transfers across machines without reshaping.
Access & license

Open by design

License

Dataset released under CC BY 4.0 for open research.

Format

One-row-per-shot Parquet with nested arrays; streamable via Hugging Face datasets.

Identifiers

Each row carries a source column — DIII-D or MAST. Machine shot numbers are not part of the release; shots are identified by their position in the config, and submissions are keyed in that same order (shot_0000_psirz, …).