gerkone/pinc_gkw

Gyrokinetics compression validation dataset (PINC)

Subsampled 5D gyrokinetics test set to reproduce the compression evaluation of Physics-Informed Neural Compression of High-Dimensional Plasma Data (PINC).

• Paper: https://arxiv.org/abs/2602.04758v2
• Code: https://github.com/ml-jku/neural-gyrokinetics
• Blog: https://ml-jku.github.io/blog/2026/pinc/

Data

60 held-out turbulent trajectories of adiabatic-electron gyrokinetic simulations (GKW), 10x-downsampled over the turbulent window (timesteps 90-260, step 10) to 1080 snapshots, stored in float32. Layout:

iteration_<n>_ifft_realpotens/
  data/timestep_<t>.bin   # 5D distribution function f at timestep t
  data/poten_<t>.bin      # electrostatic potentials at timestep t
  metadata_light.pkl, data_source.txt   # geometry, grid, spectra metadata
checkpoints/nf/
  best_mlp_*.pt      # best basic neural field, one per snapshot
  best_int_mlp_*.pt  # best PINC (physics-informed) neural field, one per snapshot

The complete trajectory set (all trajectories, bfloat16) is at gerkone/gyrokinetic-adiabatic-256traj.

Usage with PINC

git clone https://github.com/ml-jku/neural-gyrokinetics && cd neural-gyrokinetics

from huggingface_hub import snapshot_download
snapshot_download('gerkone/pinc_gkw', repo_type='dataset', local_dir='data/pinc_gkw')

Reproduce the compression evaluation (neural fields + traditional baselines) through the shared metrics pipeline:

python scripts/run_eval1k.py --methods nf,nf-pinc,sz3,jpeg2000,zfp \
    --ckpts data/pinc_gkw/checkpoints/nf --path data/pinc_gkw

See the repository README for the full PINC training and evaluation pipeline.