pedrocurvo's picture
Update README.md
3becce9 verified
metadata
license: mit
tags:
  - physics
  - fusion
  - ml
  - ai4science
pretty_name: Near-Axis Stellarators Dataset
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: stels
        path: stels/*.csv
      - split: good
        path: good-stels.csv
      - split: viable
        path: viable-stels.csv

NearAxisStellarators

The NearAxisStellarators dataset contains stellarator configurations generated with the pyQSC near-axis expansion code.
It provides input design parameters (magnetic axis Fourier coefficients, field strength coefficients, number of field periods, pressure) and the resulting plasma properties (rotational transform, elongation, Mercier stability, quasisymmetry, etc.).

This dataset supports both forward modeling (parameters → properties) and inverse design (desired properties → candidate parameters).


📊 Applications

  • Plasma physics research – explore distributions of stable stellarators.
  • Machine learning – benchmark regression, density estimation, or inverse design models.
  • Fusion design optimization – generate candidate stellarators with desired confinement properties.

🛠️ Dataset Generation

  1. Sampled design parameters from uniform distributions in physically relevant ranges.
  2. Evaluated equilibrium properties with pyQSC near-axis expansion.
  3. Applied physical constraints (Mercier stability, finite β, elongation, quasisymmetry).
  4. Iteratively retrained a Mixture Density Network (MDN) to enrich the dataset with “good” stellarators, increasing success rate from ~0.002% (random) to ~20%.

For details, see the paper and the code:


📖 Citation

If you use this dataset, please cite:

@article{Curvo2025,
          title={Using deep learning to design high aspect ratio fusion devices},
          volume={91}, DOI={10.1017/S002237782400165X}, number={1},
          journal={Journal of Plasma Physics},
          author={Curvo, P. and Ferreira, D.R. and Jorge, R.},
          year={2025},
          pages={E38}}