gerkone/pinc_gkw

Gyrokinetics compression test set (PINC)

Held-out 5D gyrokinetics test set, neural field and autoencoder checkpoints to reproduce the compression evaluation in Physics-Informed Neural Compression of High-Dimensional Plasma Data (PINC). Links to the paper, code, and authors are withheld for double-blind review; see the anonymized submission for details.

Data

60 held-out turbulent trajectories of adiabatic-electron gyrokinetic simulations (GKW), stored in float32:

• steady state: timesteps 90-260 (step 10) of every trajectory, 1080 snapshots (main results table);
• transition: 10 snapshots across the transition window of 47 trajectories, 470 snapshots (temporal-consistency table); windows and timesteps in transition_windows.json.

The electrostatic potentials are the field solve of the stored distribution function, and the per-trajectory potential statistics in metadata_light.pkl are computed from them.

iteration_<n>_ifft_realpotens/
  data/timestep_<t>.bin   # 5D distribution function f at timestep t
  data/poten_<t>.bin      # electrostatic potential at timestep t
  metadata_light.pkl      # geometry, grid, spectra metadata
transition_windows.json   # transition window (t0, t1, dt) and evaluated timesteps per trajectory
diff_df_flux_phi_offset80_mu_d0253e25_stats.pkl  # training-set normalization statistics of the autoencoders
checkpoints/
  autoencoders/           # model.eqx (equinox, weights only) + config.yaml per model
    ae/                   #   AE (1005x)
    pinc_ae/              #   PINC-AE: AE weights with the LoRA adapters
    vqvae/                #   VQ-VAE (77,369x)
    pinc_vqvae/           #   PINC-VQ-VAE: VQ-VAE weights with the LoRA adapters
    pinc_ae_single_stage/ #   single-stage PINC-AE (appendix)
  nf/                     # steady-state neural fields, one per snapshot
    best_mlp_*.pt         #   NF
    best_int_mlp_*.pt     #   PINC-NF
  nf_transition/          # transition-window neural fields (same naming)
  pinn/best_int_mlp_*.pt  # PINN baseline neural fields (steady state)

Autoencoders read the normalization statistics from the data directory, so keep the .pkl next to the trajectories. Neural field file names encode trajectory, timestep and compression ratio: <prefix>_<trajectory>_t<timestep>_x<compression ratio>.pt.

Tables (dataset viewer)

• index/snapshots.csv: one row per stored snapshot (1080 steady + 470 transition): trajectory, timestep (file index), time (R/V_r), phase (steady / transition), the physical parameters ion_temp_grad, density_grad, s_hat, q, the heat_flux at that time, the paths of the df_file and potential_file, and of the nf_checkpoint, pinc_nf_checkpoint and pinn_checkpoint (steady state only) for that snapshot.
• index/trajectories.csv: one row per trajectory (60): physical parameters, number of saved timesteps, steady snapshots and their mean heat flux, and the transition window (transition_t0, transition_t1).

Usage

Download the snapshots and checkpoints and run the evaluation script of the anonymized code submission, e.g.

python scripts/eval_compression.py --methods nf,nf-pinc,pinn,zfp,sz3,wavelet,jpeg2000 \
    --ae AE=<local>/checkpoints/autoencoders/ae:model.eqx \
    --ae PINC-AE=<local>/checkpoints/autoencoders/pinc_ae:model.eqx \
    --ae VQ-VAE=<local>/checkpoints/autoencoders/vqvae:model.eqx \
    --ae PINC-VQ-VAE=<local>/checkpoints/autoencoders/pinc_vqvae:model.eqx \
    --ckpts <local>/checkpoints/nf --pinn-ckpts <local>/checkpoints/pinn --path <local> --outdir <out>

Changelog

• 2026-10: potentials replaced by the field solve of the stored f (the previous ones were shifted), potential statistics updated; neural field checkpoints updated to the reported models; added the transition snapshots and the PINN and transition neural fields; snapshot and trajectory index tables; added the autoencoder checkpoints and their normalization statistics; PINN checkpoints now selected on PSNR(phi), as for PINC-NF.