Database overview

The BoSS holds 179,049 stellarator equilibria gathered from many sources. Every device goes through the same pipeline: an ideal MHD equilibrium is computed, followed by fast-ion confinement simulations of alpha particles.

179,049 configurations

2.7 billion alpha particles traced

134 million full particle orbits

3.2 TB of HDF5 files

Sources

The field source_dataset records where each configuration came from. The configurations come from about 97,000 distinct source devices (source_id). A boundary can appear more than once, with different pressure and current profiles, and scaled to a different machine size and field strength.

Source (source_dataset)	Description	Configurations
quasr	Quasi-axisymmetric and quasi-helically symmetric stellarators from QUASR (Giuliani et al.).	70,764
constellaration	Quasi-isodynamic-like boundaries from the ConStellaration dataset (Cadena et al., 2025).	66,135
model_improvement_2026	Boundaries generated randomly using active learning. Generated for the BoSS paper.	13,922
random_ellipse	Unoptimized rotating ellipses with random elongation, aspect ratio and axis torsion. Generated for the BoSS.	7,932
model_improvement_bo_2026	Boundaries chosen with active learning using surrogate optimization.	7,635
random_ellipse_perturbed	Rotating ellipses with random perturbations in their higher Fourier modes. Generated for the BoSS.	6,872
fast_ion_opt_2026	Boundaries optimized for alpha-particle confinement using a surrogate model. Generated for the BoSS.	5,530
omnigenity	Omnigenous equilibria computed with DESC, from the omnigenity database of Gaur et al. (2023).	185
kappel_2024	Stellarator and tokamak configurations collected by Kappel et al. (2024).	64
Others	bindel_2023: configurations optimized for fast-ion confinement by Bindel et al. (2023). bindel_2023_perturbed, quasr_perturbed: perturbed copies of a few devices.	10

Distributions

The top row shows the loss fraction of the particles born on \(s = 0.25\), the quasi-symmetry error (the square root of the summed squares of the Boozer \(\lvert B \rvert\) harmonics that break quasi-symmetry, relative to the symmetric ones, averaged over the flux surfaces) and the aspect ratio. The bottom row shows the mean rotational transform, the plasma \(\beta\) of the non-vacuum devices, and the number of field periods.

Many devices deviate strongly from quasi-symmetry, although a large share come from QUASR, a database of devices optimized for quasi-symmetry. About a third of the devices have finite pressure, with \(\langle\beta\rangle\) up to about 6%; most of them are above the reactor-grade value of 1%.

Files

The data comes in two kinds of HDF5 files: confinement files and trajectory files. They differ only in the type of guiding-center information they carry. Confinement files have particle initial and final states, whereas trajectory files contain full trajectories.

	Confinement files	Trajectory files
Fast-ion data	Start and end state of all 15,000 particles	Full orbits of 6,000 particles
Configurations per file	up to 3,000	up to 40
Train split	52 files, 279 GB, 154,128 configurations	479 files, 2.46 TB, 19,123 configurations
Test split	9 files, 45 GB, 24,921 configurations	80 files, 411 GB, 3,184 configurations

The confinement files (325 GB in all) are enough to compute loss fractions and heat fluxes. The trajectory files (2.9 TB) hold 134 million orbits for 22,307 of the configurations. See Download Data to download the data.

Train/test split

The file names indicate the split they belong to, train or test. The test split holds 24,921 configurations (14%), drawn at random across sources. All configurations that come from the same source device are in the same split, so a model is never tested on a rescaled copy of a device it was trained on.