--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - drilling - drilling-parameters - mwd - lwd - rop-optimization - mse - bit-wear - drilling-dysfunctions - xpertsystems pretty_name: "OIL-007 — Synthetic Drilling Parameters Dataset (Sample)" size_categories: - 100K MSE = (WOB / A_bit) + (120·π·RPM·T_bit) / (A_bit · ROP) where `A_bit` varies by hole section (17.5" surface, 12.25" intermediate, 8.5" production, 6.125" lateral), and `T_bit` is downhole bit torque estimated as 25-45% of surface torque depending on trajectory and depth — a critical detail for accurate MSE in long-reach and horizontal wells. **Formation strength profiles** (`mse_log.formation_strength_psi`) follow basin-specific UCS bands with stochastic formation transitions every 500-2000 ft: Permian 8-22 kpsi, Marcellus-Utica 11-25 kpsi, Brazil Pre-Salt 15-35 kpsi, Geothermal Basins 18-40 kpsi. **Rheology coupling** — PV and YP share an underlying rheology level correlated with mud weight, producing the field-realistic PV-YP correlation of ~0.70-0.85 (Bourgoyne et al. 1986). **ECD margin** (`ecd_ppg - mud_weight_ppg`) is calibrated to maintain the ~0.45 ppg overpressure circulation envelope per IADC guidance, with depth-scaled annular friction and ROP-dependent cuttings loading. **Bit wear** uses the **IADC dull grading taxonomy** (8 codes spanning worn-teeth, broken-teeth, lost-teeth, ring-out, balled-up patterns), with dull grade escalation driven by cumulative footage and average formation strength per bit run. --- ## Suggested use cases 1. **ROP optimization regression** — predict ROP from WOB, RPM, torque, mud weight, formation UCS using the 109,870-row time-series spine. 2. **MSE efficiency classification** — train models on `bit_efficiency` to identify low-energy-efficient drilling sections. 3. **Dysfunction detection** — multi-class classifier on `dysfunction_class` (11-class: stick-slip, lateral vibration, stuck pipe, washout, twist-off, mud loss/gain, bit balling, pack-off, whirl, axial bounce) from vibration + mechanical telemetry. 4. **Stuck pipe early warning** — binary classification with the `stuck_pipe_precursor` events as positive labels, ROP/WOB/torque time- series as features. 5. **Bit dull grade prediction** — regress IADC dull grade from cumulative footage and formation strength exposure per bit run. 6. **Drilling efficiency grading** — multi-class classification on the `drilling_efficiency_grade` (A+/A/B/C/D) target from per-well aggregated features. 7. **Time-series forecasting** — predict next-stand ROP / MSE / vibration from the prior stand's drilling parameters (sequence models, transformers). 8. **Multi-table relational ML** — entity-resolution and graph-based learning across the 9 joinable tables via `well_id`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil007-sample", data_files="drilling_timeseries.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd wells = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/wells_master.csv") ts = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/drilling_timeseries.csv") mse = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/mse_log.csv") events = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/drilling_events.csv") joined = ts.merge(wells, on="well_id") ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter through per-well `SeedSequence([master_seed, well_idx])` derivation, guaranteeing schema-stable joins across runs and seed-by-seed reproducibility. A seed sweep across `[42, 7, 123, 2024, 99, 1]` confirms Grade A+ on every seed in this sample. --- ## Honest disclosure of sample-scale limitations This is a **sample** product calibrated for ML prototyping and drilling- parameter research, not for live drilling operations. A few notes: 1. **Time-series is decimated 4× in the sample.** Full product runs at the generator's native cadence (~42 samples per stand at 1 Hz / 85 ft/hr). The sample uses `--decimate 4` to keep file sizes <15 MB while preserving stand-resolution detail. For high-frequency vibration ML, use the full product (12,000 wells, undecimated). 2. **ECD margin runs slightly below the 0.45 ppg target** (observed mean ~0.37 ppg) at sample scale, well within the ±0.15 ppg tolerance. This reflects the stochastic depth-scaled annular friction modeling; full- product scale converges closer to 0.45 ppg as the basin/well-type mix averages out. 3. **Max depth capped at 12,000 ft in the sample** (vs 22,000 ft in the full product). This caps the Pre-Salt and HPHT extreme-depth physics slightly under-represented. Wells assigned to those types still get their depth/mud-weight modifiers applied within the 12 kft envelope. 4. **Inter-stand jitter uses per-stand-seeded RNG** for intra-stand sample variation. This produces stand-coherent telemetry (good for ML) but means within-stand noise is correlated stand-to-stand at fixed offsets. For pure-noise modeling, filter out the per-stand structure first. 5. **Per-well stuck pipe / washout / dysfunction event rates are per-stand Bernoulli probabilities**, so total event counts per well are Poisson-distributed (~6 events/well average at sample scale). Production runs (longer wells, more stands) will see proportionally more events per well. --- ## Full product The **full OIL-007 dataset** ships at **12,000 wells**, **22,000 ft max depth**, undecimated 1Hz timeseries, full 12-class well-type coverage with HPHT and Pre-Salt extreme-depth physics, and full per-stand dysfunction event modeling — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil007_sample_2026, title = {OIL-007: Synthetic Drilling Parameters Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil007-sample} } ``` ## Generation details - Generator version : 1.0.0 - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-21 22:54:12 UTC - Wells : 100 - Max depth : 12,000 ft (capped for sample; full product: 22,000 ft) - Timeseries decim. : 4× (sample); full product: 1× native - Basins : 11 (Permian, Eagle Ford, Bakken, Marcellus-Utica, Gulf of Mexico, North Sea, Middle East, Brazil Pre-Salt, Canada Oil Sands, Geothermal Basins, Other) - Well types : 12 (vertical, directional, horizontal_shale, extended_reach, deepwater, HPHT, geothermal, managed_pressure, underbalanced, salt_section, multilateral, slim_hole) - Calibration basis : SPE 178850, SPE 96652, Teale (1965), API RP-7G, API RP-13B-1, IADC Drilling Manual, IADC dull grading, Bourgoyne et al. (1986), Rystad Energy, Spears & Associates - Overall validation: 100.0/100 — Grade A+