wellbench — Synthetic Well-Log Benchmark for Pore-Pressure Prediction

Companion dataset for the wellbench Python library: a physics-based and CTGAN-based synthetic-data benchmark calibrated against real wells from the Eastern Potwar Basin (Pakistan), released for reproducible pore-pressure prediction research.

Modality	Tabular (single-table well logs)
Generators	Physics-based (deterministic) + CTGAN baseline
Wells	9 real wells × 2 generators = 18 synthetic CSVs
Total size	~43 MB
License	MIT
Library	wellbench on PyPI
Code	https://github.com/monteirot/wellbench
Croissant	https://huggingface.co/api/datasets/monteirot/wellbench/croissant

---

Quick start

Load with 🤗 Datasets

from datasets import load_dataset

# All three Missa Keswal wells (physics generator), zone 1
ds = load_dataset("monteirot/wellbench", "physics_zone_1")
print(ds)
# DatasetDict({
#   missa_keswal_01: Dataset({features: [...], num_rows: ...}),
#   missa_keswal_02: Dataset({features: [...], num_rows: ...}),
#   missa_keswal_03: Dataset({features: [...], num_rows: ...}),
# })

# A single well as a pandas DataFrame
df = load_dataset(
    "monteirot/wellbench",
    "physics_zone_1",
    split="missa_keswal_01",
).to_pandas()
print(df.head())

Load via Croissant + mlcroissant

import mlcroissant as mlc
import pandas as pd

ds = mlc.Dataset("https://huggingface.co/api/datasets/monteirot/wellbench/croissant")
df = pd.DataFrame(ds.records(record_set="physics-samples"))

# A reproducible train / val / test split inside zone 1
train = df[df.well == "MISSA-KESWAL-01"]
val   = df[df.well == "MISSA-KESWAL-02"]
test  = df[df.well == "MISSA-KESWAL-03"]

Load directly with pandas

import pandas as pd

URL = "https://huggingface.co/datasets/monteirot/wellbench/resolve/main"
df = pd.read_csv(f"{URL}/synthetic_datasets/zone_1/synth_MISSA-KESWAL-01.csv")

Validate the Croissant metadata locally

pip install "mlcroissant[parquet]"
mlcroissant validate \
  --jsonld https://huggingface.co/api/datasets/monteirot/wellbench/croissant

---

Dataset description

wellbench is a synthetic well-log benchmark designed for reproducible methods research on pore-pressure prediction and related petrophysical inverse problems. It is the data artefact for the wellbench Python package and the arXiv:XXXX.XXXXX paper (placeholder — replace with the canonical reference once available).

The dataset addresses two long-standing pain points in machine learning for subsurface geoscience:

1. Real well logs are scarce, fragmented, and frequently encumbered by commercial confidentiality, which limits reproducible benchmarking across studies.
2. General-purpose tabular synthesisers (CTGAN, TVAE, TabDDPM, …) treat well logs as i.i.d. rows and ignore the petrophysical relationships (Athy compaction, Wyllie / Archie / Eaton equations) that any plausible log must satisfy.

wellbench provides synthetic copies of nine real wells, each generated by two complementary methods:

• A deterministic physics generator whose outputs are valid by construction — they obey the same closed-form transforms a petrophysicist would apply, with regional parameters tuned via Optuna using Jensen–Shannon divergence and Wasserstein-1 distance against real-data marginals.
• A CTGAN baseline (Xu et al. 2019) trained on the same cleaned real wells, exposed under an identical .generate() interface for side-by-side comparison.

Supported tasks

• Tabular regression — predict pore pressure (PPP) from GR, DT, RHOB, RT, depth, and the derived hydrostatic / overburden / NCT columns.
• Time-series forecasting — depth-indexed log forecasting, where DEPTH plays the role of the time axis.
• Synthetic-data evaluation — fidelity / utility / privacy comparison between physics-based and learned tabular synthesis.
• Data augmentation — pretrain or augment small real-well training sets for downstream petrophysical tasks.

Languages

Column names and documentation are in English (en). The data itself is numerical and language-agnostic.

---

Dataset structure

File layout

monteirot/wellbench
├── synthetic_datasets/                  # Physics-based generator outputs
│   ├── zone_1/                          # Missa Keswal (Eastern Potwar Basin)
│   │   ├── synth_MISSA-KESWAL-01.csv
│   │   ├── synth_MISSA-KESWAL-02.csv
│   │   └── synth_MISSA-KESWAL-03.csv
│   ├── zone_2/                          # Pindori field
│   │   ├── synth_PINDORI-1.csv
│   │   ├── synth_PINDORI-2.csv
│   │   └── synth_PINDORI-3.csv
│   └── zone_3/                          # Joyamair / Minwal area
│       ├── synth_JOYAMAIR-4.csv
│       ├── synth_MINWAL-2.csv
│       └── synth_MINWAL-X-1.csv
├── CTGAN_synthetic_data/                # CTGAN-baseline outputs (same layout)
│   ├── zone_1/  …
│   ├── zone_2/  …
│   └── zone_3/  …
├── README.md                            # this dataset card
├── croissant.json                       # MLCommons Croissant 1.0 metadata
└── .gitattributes

Configurations and splits

The dataset exposes 6 configs (2 generators × 3 zones), with one split per real well under each config:

Config	Generator	Zone	Splits	Default
physics_zone_1	Physics	1	missa_keswal_01 · missa_keswal_02 · missa_keswal_03	✅
physics_zone_2	Physics	2	pindori_1 · pindori_2 · pindori_3
physics_zone_3	Physics	3	joyamair_4 · minwal_2 · minwal_x_1
ctgan_zone_1	CTGAN	1	missa_keswal_01 · missa_keswal_02 · missa_keswal_03
ctgan_zone_2	CTGAN	2	pindori_1 · pindori_2 · pindori_3
ctgan_zone_3	CTGAN	3	joyamair_4 · minwal_2 · minwal_x_1

A suggested train/val/test mapping mirroring the convention used in the paper:

Zone	train	validation	test
1	missa_keswal_01	missa_keswal_02	missa_keswal_03
2	pindori_1	pindori_2	pindori_3
3	joyamair_4	minwal_2	minwal_x_1

Users are free to ignore this convention; each split is a self-contained synthetic well.

Data instances

Each row is a single depth measurement of a single (synthetic) well. An example row from physics_zone_1 / missa_keswal_01:

{
  "DEPTH": 500.0,
  "GR": 140.2155,
  "DT": 138.1432,
  "RHOB": 1.6418,
  "RT": 12.7382,
  "HP": 645.94,
  "OB": 926.85,
  "DT_NCT": 137.43,
  "PPP": 615.33
}

Data fields

All 18 CSVs in this release belong to pore-pressure regions (zones 1–3), so all of the columns below are present:

Column	Type	Units / range	Description
DEPTH	float	ft (region-dependent)	Measured depth — the row index along the well axis
GR	float	API, [0, 200]	Gamma ray (shale-volume proxy)
DT	float	µs/ft, [30, 180]	Sonic transit time
RHOB	float	g/cc, [1.2, 2.9]	Bulk density
RT	float	Ω·m, [0.01, 10 000]	True resistivity
HP	float	psi	Hydrostatic pressure
OB	float	psi	Overburden pressure
DT_NCT	float	µs/ft	Sonic normal-compaction-trend reference
PPP	float	psi	Pore pressure (Eaton's method) — primary target

All values are clipped to wellbench.PHYSICAL_BOUNDS, the same physical envelope enforced by the library's clean_well_data routine.

---

Dataset creation

Curation rationale

Public well-log corpora suitable for ML benchmarking are dominated by a handful of repeatedly-cited datasets (e.g. Volve, Force-2020), whose pore-pressure labels are not always disclosed and whose proprietary status complicates redistribution. wellbench provides a fully synthetic, openly licensed alternative that:

• Is calibrated to real-world basin parameters rather than sampled from a synthetic prior in vacuum.
• Pairs physics and learned generators under one schema, so studies can isolate the contribution of physical priors.
• Ships under MIT so there is no friction for downstream redistribution.

Source data

Regional parameter sets (REGION_1, REGION_2, REGION_3 in the wellbench library) were calibrated against publicly described wells from the Eastern Potwar Basin, Punjab, Pakistan:

• Zone 1 — Missa Keswal field (MISSA-KESWAL-01, -02, -03)
• Zone 2 — Pindori field (PINDORI-1, -2, -3)
• Zone 3 — Joyamair / Minwal area (JOYAMAIR-4, MINWAL-2, MINWAL-X-1)

The synthetic outputs published here are not anonymised real data; they are generated outputs of the calibrated wellbench models. Real-well log values are not redistributed.

Generation pipeline

Physics generator

1. Athy's exponential compaction → porosity profile φ(z).
2. Wyllie time-average equation → DT from φ.
3. Bulk-density mixing law → RHOB from φ + lithology trend.
4. Archie's equation → RT from φ + saturation prior.
5. Shale-volume linear mixing → GR.
6. Hydrostatic + overburden integration; Eaton's method on the normal compaction trend → HP, OB, DT_NCT, PPP.

Per-region parameters were optimised with Optuna to minimise a weighted sum of Jensen–Shannon divergence and Wasserstein-1 distance between synthetic and real marginals across all log columns. The library version, region dictionary, and seed used to produce each CSV are recorded in croissant.json.

CTGAN baseline

The five CTGAN checkpoints bundled in the wellbench package (ctgan_r1.pkl … ctgan_r5.pkl) were trained on cleaned real wells from each region. CTGAN samples are i.i.d. rows; wellbench orders them along a user-supplied depth axis, applies a column-rename to match the physics schema, and clips to PHYSICAL_BOUNDS. The CSVs in CTGAN_synthetic_data/ were generated with deterministic seeds and the bundled checkpoints — see wellbench.load_ctgan_generator(region_index=…) for the exact loader.

Annotations

None. The dataset is fully machine-generated; there is no human annotation step.

Personal and sensitive information

The synthetic data does not contain personally identifying information. Real wells used for calibration are publicly described in the petroleum-geology literature; no proprietary log values are redistributed in this release.

---

Considerations for using the data

Social impact

Open synthetic well logs lower the barrier to entry for ML research on subsurface geoscience and reduce reliance on confidentiality-restricted operator data. Better pore-pressure prediction tools have direct safety implications for drilling.

Discussion of biases

• Geographic bias. All nine wells are from a single basin (Eastern Potwar). Generalisation to other tectonic settings is the user's responsibility.
• Lithological bias. The bundled regions emphasise shale-dominated, clastic settings. Carbonates and evaporites are not represented.
• Calibration target bias. Calibration optimises marginal fidelity (JS + Wasserstein-1) — joint fidelity and cross-column residual structure are not directly optimised, so multivariate downstream models may see slightly different correlation structure than in real wells.
• Public-record selection bias. Real wells used for CTGAN training are those whose logs are publicly described. Wells absent from the open record (often the most challenging ones) are by construction absent here.

Known limitations

• Synthetic ≠ real. Subtle stratigraphic features useful for prospect evaluation are deliberately not modelled. Use this dataset for methods research, not for operational prospect ranking.
• CTGAN depth ordering is post-hoc. CTGAN samples i.i.d. rows; vertical correlation is therefore weaker than in the physics generator and weaker than in real wells. This is by design — it is a baseline for comparison.
• Eaton-only PP model. Pore-pressure outputs reflect under-compaction signatures; other overpressure mechanisms (kerogen maturation, tectonic loading) are not modelled.
• Three zones only. The library defines five regions, but only the three pore-pressure regions are included in this release.

Out-of-scope use

• ❌ Operational drilling decisions. The synthetic logs capture distributional properties of real basins, not the physics of any specific surveyed location.
• ❌ Substituting for proprietary well data. Use this for benchmarking, not to evade data-sharing constraints.
• ❌ Privacy proxy without further audit. The physics generator does not memorise real samples by construction, but the bundled CTGAN baseline was trained on cleaned real wells and has not been audited for membership inference. Treat CTGAN outputs as research artefacts, not as anonymised data.
• ❌ Training models that will be deployed against carbonate or evaporite reservoirs without recalibration.

Recommended evaluation

When using this dataset to benchmark pore-pressure or petrophysical models, we recommend reporting at minimum:

• Fidelity. Jensen–Shannon divergence and Wasserstein-1 distance per log column, vs. the corresponding real well distribution (where available to the user).
• Utility. TSTR (train-on-synthetic, test-on-real) AUROC / RMSE on a downstream pore-pressure classifier or regressor.
• Generator comparison. Run the same downstream model on physics_zone_* and ctgan_zone_* and report the gap.

A reference evaluation harness is provided in wellbench.metrics.

---

Additional information

Dataset curators

Maintained by @monteirot. See the GitHub repository for the full author list and affiliations.

Licensing information

This dataset is released under the MIT License. The synthetic outputs are derivative of the wellbench library, which is also MIT-licensed. Real wells used for calibration are publicly described in the petroleum-geology literature and their values are not redistributed here.

Citation

If you use wellbench data in your research, please cite both the paper and the software release:

@inproceedings{wellbench2026,
  title     = {wellbench: A Physics-Calibrated Synthetic Benchmark for Pore-Pressure Prediction},
  author    = {Monteiro, T. and ...},
  booktitle = {Advances in Neural Information Processing Systems, Datasets and Benchmarks Track},
  year      = {2026},
  url       = {https://arxiv.org/abs/XXXX.XXXXX}
}

@software{wellbench_software,
  author  = {Monteiro, T. and ...},
  title   = {wellbench: Physics-based synthetic well-log benchmark generator},
  year    = {2026},
  version = {<release>},
  doi     = {10.5281/zenodo.XXXXXXX},
  url     = {https://github.com/monteirot/wellbench}
}

For the dataset specifically, please also link to https://huggingface.co/datasets/monteirot/wellbench so the Croissant record is discoverable.

Contributions

Issues, corrections, and additional regional calibrations are very welcome. Please open a discussion on the Hugging Face Community tab or a pull request on GitHub.

Reproducing the dataset

The CSVs in this repository are exactly reproducible from the wellbench library:

pip install "wellbench[ctgan]"

# Physics generator — zone 1, all three Missa Keswal wells
python -c "
import pandas as pd
from wellbench import SyntheticWellLogGenerator, REGION_1
real = pd.read_csv('real_well.csv', usecols=['DEPTH'])  # supply your own
SyntheticWellLogGenerator(REGION_1).generate(
    seed=42, depth=real['DEPTH'].to_numpy(),
).to_csv('synth_MISSA-KESWAL-01.csv', index=False)
"

# Or reproduce the canonical 15-dataset benchmark
wellbench -o benchmark/

The exact commands and seeds used to produce each CSV are recorded in the wellbench repository under benchmarks/ and in the Croissant metadata (croissant.json) of this dataset.