MHD equilibrium surrogates of a quasi-isodynamic stellarator
Neural-network surrogates that map a pressure profile to the fixed-boundary DESC
equilibrium of a quasi-isodynamic stellarator configuration (NFP = 3, stellarator
symmetric, scaled to B0 = 1 T): the Fourier-Zernike coefficients R_lmn, Z_lmn and
lambda_lmn, the rotational transform iotaf and the toroidal flux Psi. They were
trained for the Master's thesis Bayesian Experimental Design of Fusion Plasma
Diagnostics with Self-Consistent MHD Equilibria (Technical University of Munich, 2026).
- Training data: proxima-fusion/diagnostics-bayesian-experimental-design-datasets
- Code: diagnostics_bayesian_experimental_design
- License: CC BY 4.0
| Model | Training |
|---|---|
20260907_081512_mlp_data_fnorm_only_pca_hand_tuned |
Data loss for 1101 epochs, then the DESC force residual alone for 2000; the thesis code's default surrogate |
20260908_141134_mlp_data_pca_hand_tuned |
Data loss alone, 1101 epochs |
Both are MLPs (Flax, 3 hidden layers of 512, leaky ReLU, 1,214,684 parameters) from the
pressure at the 99 spline knots to 1244 PCA coefficients of the 2955-long target
[R_lmn | Z_lmn | lambda_lmn | iotaf | Psi]. Each model is two files: <name>.npz holds
the weights, the preprocessing and the basis, and <name>.json the architecture,
training curriculum and training record. models_training/<name>/ holds the per-epoch
training history and the per-stage summary.
Usage
The code repository downloads these files, and the datasets, into place with its
download script (see its README), and loads a model with
model_storage_utils.load_mlp_model.
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}
}