proxima-fusion/diagnostics-bayesian-experimental-design-datasets

Datasets of fixed-boundary DESC equilibria

Datasets of fixed-boundary DESC equilibria for a quasi-isodynamic stellarator configuration (NFP = 3, stellarator symmetric) with the corresponding pressure profiles. The configuration is scaled to B0 = 1 T. The Fourier-Zernike resolution used was L = M = 11, N = 12.

These equilibria are used as training and test sets of the MHD equilibrium surrogates of the Master's thesis Bayesian Experimental Design of Fusion Plasma Diagnostics with Self-Consistent MHD Equilibria (Technical University of Munich, 2026).

• Trained surrogates: proxima-fusion/diagnostics-bayesian-experimental-design-surrogates
• Code: diagnostics_bayesian_experimental_design
• License: CC BY 4.0

File	Content
equilibrium_a_3k_spline_input.parquet	3000 equilibria: 2400 Gaussian-process and 600 parametric p0 (1 - s^gamma)^alpha pressure profiles, p0 <= 35 kPa
equilibrium_a_interpolator_28_spline_input.parquet	28 equilibria on the lattice of p0 (1 - rho^a)^b: p0 in {0, 11.7, 23.3, 35} kPa, a in {2, 4, 6}, b in {0.5, 2.25, 4}
<dataset>_settings.json	Spectral basis, mode tables of R_lmn, Z_lmn and lambda_lmn, radial grids; the 3k file also holds the train/val/test split, the lattice file its nodes
equilibrium_a_warm_start.json	The DESC starting point every row was solved from: the vacuum equilibrium's coefficients, magnetic axis and boundary

Columns: pressure_knots and pressure_params (the pressure spline the solver saw, Pa; pressure repeats the values), Psi (Wb), R_lmn, Z_lmn, lambda_lmn, iotaf, beta_volume_averaged, the fnorm_volume_averaged_* force residual, and desc_equilibrium_id (row id).

Generation

A vacuum equilibrium was computed once with VMEC (fixed boundary), converted to DESC at L = M = 11, N = 12 and solved again with DESC. Every row is then one DESC Equilibrium.solve(objective="force"), lsq-exact (trust-region least squares) at DESC's default tolerances, started from that vacuum equilibrium with the row's pressure profile. The code repository's regenerate_dataset_row.py rebuilds that start from equilibrium_a_warm_start.json and solves any row again by its desc_equilibrium_id.

Dataset	Wall time	Machine
equilibrium_a_3k	2 h 23 min	250 cloud workers in parallel
equilibrium_a_interpolator_28	1 h 25 min (252 to 2122 s per solve)	AMD EPYC 7B12, 32 vCPU, 125 GB RAM, 7 solves in parallel

Usage

The code repository downloads these files, and the surrogates, into place with its download script (see its README). Without the code:

import pandas as pd

equilibria = pd.read_parquet(
    "hf://datasets/proxima-fusion/diagnostics-bayesian-experimental-design-datasets/"
    "equilibrium_a_3k_spline_input.parquet"
)

Citation

@mastersthesis{valente2026bayesian,
  author = {Valente, Matilde},
  title  = {Bayesian Experimental Design of Fusion Plasma Diagnostics with Self-Consistent {MHD} Equilibria},
  school = {Technical University of Munich},
  year   = {2026}
}