Upload croissant.json

croissant.json

@@ -16,6 +16,16 @@
       "@type": "@vocab"
     },
     "dct": "http://purl.org/dc/terms/",
+    "prov": "http://www.w3.org/ns/prov#",
+    "wasGeneratedBy": "prov:wasGeneratedBy",
+    "wasDerivedFrom": "prov:wasDerivedFrom",
+    "dataBiases": "rai:dataBiases",
+    "dataCollection": "rai:dataCollection",
+    "dataLimitations": "rai:dataLimitations",
+    "dataUseCases": "rai:dataUseCases",
+    "dataSocialImpact": "rai:dataSocialImpact",
+    "hasSyntheticData": "rai:hasSyntheticData",
+    "personalSensitiveInformation": "rai:personalSensitiveInformation",
     "equivalentProperty": "https://schema.org/equivalentProperty",
     "examples": {
       "@id": "cr:examples",
@@ -84,6 +94,35 @@
   "isAccessibleForFree": true,
   "isLiveDataset": false,
   "citeAs": "Monteiro, T. (2026). wellbench: synthetic well-log benchmark for pore-pressure prediction (Version 0.1.2) [Dataset]. Hugging Face. https://huggingface.co/datasets/monteirot/wellbench",
+  "hasSyntheticData": true,
+  "dataLimitations": "All 18 CSVs are synthetic. The physics partition is bound by the assumptions of the deterministic generator (1D effective-stress / Eaton model, region-specific calibration, and idealized lithology) and does not reproduce drilling artifacts, washouts, tool noise, invasion, gas effects, fractures, or salt. The CTGAN partition inherits the limitations of the physics partition (since it was trained on it) and additionally exhibits the typical failure modes of tabular GANs: depth ordering is not preserved, vertical autocorrelation between adjacent samples is weakened, and tail values can be mis-calibrated. Geographic coverage is limited to nine wells across three Eastern Potwar Basin (Pakistan) fields; results obtained on wellbench should not be extrapolated to other basins, deep-water plays, unconventional reservoirs, or HPHT environments without further validation. Sample size is small (9 wells per partition) and is intended for benchmarking and reproducibility studies, not for training production models. Wellbench is explicitly not recommended for: operational drilling decisions, real-time pore-pressure monitoring, regulatory submissions, casing-design qualification, or any safety-critical use.",
+  "dataBiases": "Selection bias: the five regional calibrations and the nine wells in this release were chosen for tractability and public availability of analog parameters, not for statistical representativeness of global subsurface conditions. Coverage is concentrated in onshore Pakistan, with smaller coverage of Bering Sea and North Sea analogs in the wider wellbench generator (not all of which are uploaded here). Generator bias: the physics engine encodes the modeller's assumptions about normal-compaction behavior and the validity of the Eaton method for pore-pressure inference; basins or formations where these assumptions fail will be under-represented. CTGAN bias: the GAN was trained on the physics partition and therefore amplifies any structural bias already present there; it does not introduce independent real-well variability. Class/range imbalance: most depth samples sit in normal-pressure intervals, so over-pressured zones are under-represented relative to their operational importance; models trained directly on the unbalanced data may underestimate high-pressure risk.",
+  "personalSensitiveInformation": "None. Wellbench contains no personal, demographic, health, financial, or operator-identifiable information. All values are synthetic. Well names (e.g. MISSA-KESWAL-01, PINDORI-2, MINWAL-X-1) refer to publicly known field names in the Eastern Potwar Basin and are used only as analog labels; no proprietary measurements, lease terms, production figures, or commercially sensitive operator data are included.",
+  "dataUseCases": "Construct: each row is intended to represent a depth sample of a synthetic well, with the field PPP standing in for the pore pressure that a downstream model would predict from the four log channels (GR, DT, RHOB, RT) and the derived quantities (HP, OB, DT_NCT). Validity is established for: (1) reproducible benchmarking of regression models for pore-pressure prediction, (2) controlled comparison of physics-based vs GAN-based tabular synthesis, (3) ablation studies where a known ground truth (Eaton-method PPP) is required, (4) educational use in petrophysics and ML courses. Validity is NOT established for: (a) transfer to real wells without domain-adaptation experiments, (b) uncertainty-quantification claims about real subsurface pressure, (c) basin-scale geomechanical inference, (d) any safety-relevant or operational drilling use.",
+  "dataSocialImpact": "Positive: lowers the barrier to entry for ML-on-subsurface research by replacing proprietary, license-restricted well-log data with an open, reproducible synthetic benchmark; enables fair leaderboard-style comparisons; supports teaching without requiring an industry NDA. Negative / risks of misuse: (a) practitioners may deploy models that perform well on wellbench against real wells, where distributional differences could lead to under-estimated pore pressure and, in the worst case, contribute to drilling incidents (kicks, blowouts); (b) synthetic-only training risks giving a false sense of model robustness; (c) results on wellbench should not be cited as evidence of real-world deployability of a method. Mitigations: the dataset card and this metadata explicitly mark all data as synthetic, document the generator's assumptions, and warn against operational use; the benchmark recommends well-level (not row-level) train/val/test splits to discourage leakage-driven optimism; the upstream wellbench generator is open-source so users can inspect and adjust the physics.",
+  "dataCollection": "No human subjects, no field measurements, no annotation. The physics partition was produced by the deterministic wellbench generator (https://github.com/tiagomonteiro0715/wellbench, v0.1.2) by sampling depth grids per region, simulating GR / DT / RHOB / RT logs from calibrated parameter ranges, then computing HP, OB, DT_NCT, and PPP analytically (Eaton method). The CTGAN partition was produced by training the optional ctgan extra of wellbench on the physics partition and sampling synthetic rows per zone. Random seeds and region configurations are encoded in the source repository (src/regions.py). All generation was run locally on the dataset author's machine; no third-party annotation service or crowdworker platform was involved.",
+  "wasDerivedFrom": [
+    {
+      "@type": "sc:CreativeWork",
+      "name": "wellbench (Python package)",
+      "url": "https://pypi.org/project/wellbench/",
+      "version": "0.1.2",
+      "description": "Open-source physics-based synthetic well-log generator. The CSVs in this dataset are direct outputs of this package's CLI."
+    },
+    {
+      "@type": "sc:CreativeWork",
+      "name": "wellbench (source repository)",
+      "url": "https://github.com/tiagomonteiro0715/wellbench",
+      "description": "Generator source code and region calibrations."
+    }
+  ],
+  "wasGeneratedBy": {
+    "@type": "prov:Activity",
+    "name": "wellbench-generation",
+    "description": "Two-stage generation pipeline. Stage 1 (physics): for each of three zones (Eastern Potwar Basin: Missa Keswal, Pindori, Joyamair/Minwal), the wellbench v0.1.2 deterministic engine simulated three wells per zone. For each well a depth grid was constructed from the region calibration; GR, DT, RHOB, and RT were sampled per depth from physically plausible distributions; hydrostatic pressure (HP) and overburden (OB) were integrated; the normal compaction trend (DT_NCT) was fit; and pore pressure (PPP) was computed via the Eaton method. Outputs were written as CSV (one file per well) under synthetic_datasets/zone_{1,2,3}/. Stage 2 (CTGAN): a Conditional Tabular GAN (ctgan extra of wellbench) was trained per zone on the corresponding physics CSVs and sampled to produce 9 baseline CSVs under CTGAN_synthetic_data/zone_{1,2,3}/. No human annotation, crowdsourcing, or external data was used. Reproducibility: rerunning the wellbench CLI at the same version with the seeds in src/regions.py reproduces the physics partition byte-identically; the CTGAN partition is reproducible up to torch's CUDA non-determinism. Generation hardware: a single workstation (no distributed compute, no managed annotation platform).",
+    "startedAtTime": "2026-04-29",
+    "endedAtTime": "2026-04-29"
+  },
   "distribution": [
     {
       "@type": "cr:FileObject",