Yapenge/EXL50U-PTEFIT

EXL-50U PTEFIT Equilibrium Dataset

Magnetic equilibrium reconstruction results for the EXL-50U tokamak, produced with PTEFIT using the real-time TensorRT engine (RTEqInv). Each file is one discharge (shot); each Parquet row stores the full time series and 2-D fields for that shot.

Access

This dataset is public but gated. To download:

1. Log in to Hugging Face.
2. Open this dataset page and click Agree and access repository (or Request access).
3. Wait for manual approval if required.
4. Download with the Hugging Face CLI or Python (see below).

Dataset summary

Item	Value
Device	EXL-50U (device = EXL-50U-V58.6)
Shots	354
Format	Parquet (one file per shot)
Inversion	RTEqInv (TensorRT), one launch per time slice, recurrent state across the shot
License	CC BY-NC 4.0 — non-commercial use

File layout

exl50u_shot_{shot}.parquet   # e.g. exl50u_shot_20701.parquet

Each file contains one row (one shot). Nested arrays hold time series and 2-D fields. The Hugging Face Dataset Viewer may not preview this split because rows contain large nested arrays (inv_psi, inv_jphi, etc.). Use the Files and versions tab or download files programmatically.

Row schema

Identity and grid

Column	Shape	Unit	Description
shot	scalar	—	Discharge number
time	(T,)	ms	Successful inversion times
R	(129,)	m	Radial grid for inv_psi
Z	(129,)	m	Vertical grid for inv_psi

Metadata

Column	Example	Description
source	PTEFIT	Software / pipeline name
device	EXL-50U-V58.6	Device configuration version
inversion_backend	RTEqInv	TensorRT single-step inversion per slice

T denotes the number of successful time slices for that shot (varies per discharge).

Diagnostic inputs (diag_*)

Measurements passed to the inversion at each time step (flux in Wb/rad; IP is measured plasma current, not the inverted IP).

Column	Shape	Unit	Description
diag_flux	(T, 47)	Wb/rad	Flux loops
diag_mpt	(T, 84)	T	Tangential magnetic probes
diag_mpn	(T, 80)	T	Normal magnetic probes
diag_IP	(T,)	A	Measured plasma current
diag_eddy_current	(T,)	A	Total eddy current (often zero)
diag_vloops	(T, 47)	V	Loop voltages (often zero)
diag_CS	(T,)	A	Central solenoid current
diag_PF	(T, 14)	A	PF coil currents (PF1–PF14)
diag_TF	(T,)	A	Toroidal field coil current

Constraint weights (weights_*)

Per-channel weights recorded with the dataset. No time dimension (one weight vector per shot).

Column	Shape	Unit	Description
weights_flux	(47,)	—	Flux loop weights
weights_mpt	(84,)	—	MPT weights
weights_mpn	(80,)	—	MPN weights
weights_IP	scalar	—	IP constraint weight (often 1e-6)
weights_eddy_current	scalar	—	Eddy constraint weight
weights_vloops	(47,)	—	Vloop weights
weights_CS	scalar	—	Recorded weight (not an LSQ constraint)
weights_PF	(14,)	—	Recorded weight (not an LSQ constraint)
weights_TF	scalar	—	Recorded weight (not an LSQ constraint)

Note: weights_* values are taken from the offline HDF5 export (_build_weights()). The RTEqInv engine uses weights baked into runtime/model.pt from TOML [eq].weights; the two may differ for some channels.

Inversion results (inv_*)

Column	Shape	Unit	Description
inv_psi	(T, 129, 129)	Wb/rad	Poloidal flux Ψ(R, Z)
inv_jphi	(T, 129, 128)	A/m²	Toroidal current density (Z has one fewer point than inv_psi)
inv_surfs	(T, 2, 90)	m	LCFS polygon; index 0 = R, 1 = Z
inv_isoflux	(T, 9)	Wb/rad	Flux at nine configured iso-flux points
inv_rmin, inv_rmax	(T,)	m	LCFS min / max R
inv_zmin, inv_zmax	(T,)	m	LCFS min / max Z
inv_rc, inv_zc	(T,)	m	LCFS geometric center
inv_a	(T,)	m	LCFS minor radius
inv_kappa	(T,)	—	Elongation
inv_deltal, inv_deltau	(T,)	—	Lower / upper triangularity
inv_psi_a, inv_psi_b	(T,)	Wb/rad	Axis / boundary flux
inv_ra, inv_za	(T,)	m	Magnetic axis coordinates
inv_rx, inv_zx	(T, 4)	m	X-point candidates (0 if invalid)
inv_li	(T,)	—	Normalized internal inductance
inv_betap, inv_betat	(T,)	—	Poloidal / toroidal β

Dimensionless quantities (inv_kappa, inv_deltal, inv_deltau, inv_li, inv_betap, inv_betat) have unit none.

Inversion settings

• Engine: RTEqInv (TensorRT deployment bundle runtime/model.pt)
• Slices: One TensorRT launch per time slice; recurrent state (jphi, psi_a, psi_b, …) carried across the shot
• Input: magfield.mat per discharge
• Skipped slices: Plasma current below 100 kA (MIN_IP_A)
• Grid: 129×129 for inv_psi; Green table greens129x129.dic

Results are not multi-epoch EqInv iterations. Do not expect numerical identity with offline EqInv HDF5 exports that use many iterations per slice.

Usage

Download one shot

export HF_ENDPOINT=https://hf-mirror.com   # if direct huggingface.co is unreachable
hf auth login
hf download Yapenge/EXL50U-PTEFIT exl50u_shot_20701.parquet --repo-type dataset

Read with pandas

import numpy as np
import pandas as pd

row = pd.read_parquet("exl50u_shot_20701.parquet").iloc[0]
shot = int(row["shot"])
time_ms = np.asarray(row["time"], dtype=np.int32)           # (T,)
psi = np.stack(row["inv_psi"])                                # (T, 129, 129)
flux = np.stack(row["diag_flux"])                             # (T, 47)
w_flux = np.asarray(row["weights_flux"], dtype=np.float32)    # (47,)
kappa = np.asarray(row["inv_kappa"], dtype=np.float32)        # (T,)

Read with huggingface_hub

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="Yapenge/EXL50U-PTEFIT",
    filename="exl50u_shot_20701.parquet",
    repo_type="dataset",
)
row = pd.read_parquet(path).iloc[0]

Select one time slice

import numpy as np

target_t = 500  # ms
idx = int(np.where(time_ms == target_t)[0][0])
psi_t = psi[idx]
rc_t = float(row["inv_rc"][idx])

Important notes

• time, diag_*, and inv_* share the same length T per shot.
• Do not reshape inv_jphi as (129, 129); its Z dimension is 128.
• weights_CS, weights_PF, and weights_TF are stored for completeness; they are not inversion constraint weights in the PTEFIT layout.
• Nested columns are stored as lists / arrays inside each Parquet row; always np.asarray() or np.stack() before numerical use.

Citation

If you use this dataset, please cite PTEFIT and acknowledge the EXL-50U team. Add your publication reference here when available.

License

This dataset is released under CC BY-NC 4.0. You may use it for non-commercial research with attribution. Commercial use is not permitted under this license.