--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - mud-logging - formation-evaluation - gas-chromatography - pixler-ratios - pore-pressure - kick-detection - lithology-id - xpertsystems pretty_name: "OIL-009 — Synthetic Mud Logging Dataset (Sample)" size_categories: - 100K wetness_ratio = (C2 + C3 + C4 + C5) / (C1 + C2 + C3 + C4 + C5) × 100 > balance_ratio = C1 / (C2 + C3) > character_ratio = (C4 + C5) / C3 These are the three ratios used by every commercial mud-logging service (Halliburton, SLB, Geoservices, Pason) to classify shows as dry gas / wet gas / condensate / oil. Sample wetness mean ~12 is in the **wet-gas / oil- rich-gas** Pixler zone (5-17.5); sample balance ~10 is in the **light oil** zone (1.5-100). **`gas_readings.csv`** — basin-specific gas means with in-reservoir amplification (1.65×) and overpressure amplification (1.28×), plus log-normal noise. Background gas levels match the IADC mud-logging convention (50-500 units typical, >500 anomalous). **`pore_pressure_indicators.csv`** — implements the **Jorden & Shirley (1966) d-exponent** overpressure detection method: > d = log(ROP/60·N) / log(12·WOB/10⁶·D) (corrected for mud weight) Normal-compaction shale baseline is ~1.0-1.5; values decreasing with depth indicate undercompacted shales and impending overpressure. Sample mean d-exponent ~1.30 with downward deviations correlating with the `overpressure_flag` column. **`lithology_intervals.csv`** — 9-class lithology (shale, sandstone, siltstone, limestone, dolomite, marl, anhydrite, volcanic, coal) drawn from basin-conditioned probability mixes. Shale dominates at 36% reflecting the modern unconventional-heavy global drilling portfolio. **`drilling_events.csv`** — 10-class event taxonomy (normal drilling, drilling break, connection gas, trip gas, lost circulation, kick precursor, tight hole, differential sticking, sensor dropout, lag correction). Kick precursors gated by (overpressure + gas > 220 + mud-weight underbalance); drilling breaks gated by (hydrocarbon show + 38% draw rate). --- ## Suggested use cases 1. **Pixler hydrocarbon-ratio classification ML** — train classifiers on wetness / balance / character ratios → dry-gas / wet-gas / condensate / oil / no-show labels. Pixler crossplot zones are well-separated targets. 2. **Lithology identification from gas + chromatograph** — multi-class classifier on `lithology_label` (9-class) from drilling mechanics + gas composition + cuttings fluorescence features. 3. **Kick-risk early warning** — binary classifier on `kick_risk_flag` from upstream features (d-exponent decline, gas elevation, mud-weight underbalance). Sample has 1% positive rate matching IADC field statistics. 4. **Pore-pressure regression** — regress `pore_pressure_ppg_equiv` from d-exponent, shale density, depth, and drilling-mechanics features. 5. **Hydrocarbon show detection** — binary classifier on `hydrocarbon_show_flag` from gas + fluorescence + lithology features. 6. **Reservoir quality grading** — multi-class classifier on `reservoir_quality` (low/medium/high) from petrophysical and show-related features. 7. **Drilling event classification** — 10-class classifier on `drilling_event_type` from time-series drilling-mechanics features. 8. **Multi-table relational ML** — entity-resolution and graph-based learning across the 12 joinable tables via `well_id` and depth. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil009-sample", data_files="gas_readings.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd gas = pd.read_csv("hf://datasets/xpertsystems/oil009-sample/gas_readings.csv") chr_ = pd.read_csv("hf://datasets/xpertsystems/oil009-sample/gas_chromatography.csv") lith = pd.read_csv("hf://datasets/xpertsystems/oil009-sample/lithology_intervals.csv") lbl = pd.read_csv("hf://datasets/xpertsystems/oil009-sample/drilling_labels.csv") joined = gas.merge(chr_, on=["well_id","depth_ft"]).merge(lbl, on=["well_id","depth_ft"]) ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving `random.Random(seed)`). 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 mud-logging research, not for live drilling decisions. A few notes: 1. **Long-tail lithology classes are under-represented at sample scale.** Anhydrite (~0.9%), volcanic (~1.4%), and coal (~0.3%) are rare classes that appear only when their parent basins are drawn. Full product (18,000 wells) gives sufficient samples for these rare classes; the sample provides only handful-of-rows demonstrations of the schema. 2. **All detail tables are co-resolved at the depth_step granularity** (100 ft in the sample). Real mud-logging data has higher-frequency gas readings (1-5 ft intervals) and lower-frequency cuttings descriptions (5-30 ft intervals). The schema is the same; only the resolution differs. For high-frequency ML, use the full product with `--depth-step-ft 5`. 3. **Anomaly injection rate is 3%** (`anomaly_injection_rate=0.03`) — gas units randomly multiplied by [0.25, 0.45, 1.9, 2.8] to simulate sensor dropouts and lag corrections. These appear as outliers in `gas_readings` and can be filtered out via `gas_chromatography.gas_quality_flag == 1`. 4. **Hydrocarbon show rate (10%) and kick risk rate (1%) match aggregate IADC field statistics** but are not stratified by basin. Per-basin show rates in real data range from 2-3% (Marcellus dry gas) to 25-30% (Pre-Salt carbonate plays). Future generator v1.1 will introduce basin-conditioned show priors. 5. **Mud-log timeseries uses `mud_log_timeseries.csv`** as the canonical time-axis spine — all other tables (gas, lithology, cuttings, etc.) are indexed at the same depth grid for clean ML joins. This makes the tables more relational and less "time-series-y" than real MWD/LWD streams; treat the sample as **depth-domain mud-log records**, not time-domain telemetry. --- ## Full product The **full OIL-009 dataset** ships at **18,000 wells** with **~9M depth- resolved mud-log records**, 5-ft default depth resolution, basin-conditioned hydrocarbon show priors, and per-basin chromatograph stratification — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil009_sample_2026, title = {OIL-009: Synthetic Mud Logging Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil009-sample} } ``` ## Generation details - Generator version : 1.0-file1-simulation-engine - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-21 23:20:37 UTC - Wells : 150 - Depth step : 100 ft - Anomaly rate : 3.0% - Basins : 10 (Permian, Eagle Ford, Bakken, Marcellus, GoM Deepwater, North Sea, Middle East Carbonate, Brazil Pre-Salt, Canadian Oil Sands, Tight Gas Sandstone) - Lithologies : 9 (shale, sandstone, siltstone, limestone, dolomite, marl, anhydrite, volcanic, coal) - Calibration basis : Pixler (1969), Jorden & Shirley (1966), IADC Mud Logging Standards, IADC Well Control Statistics, API RP-13B-1, SPE 142884, Schlumberger Mud Logging Field Manual, Halliburton Mud Logging guide - Overall validation: 100.0/100 — Grade A+