--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - well-control - kick-detection - blowout-prevention - bop-operations - safety-engineering - iadc - xpertsystems pretty_name: "OIL-011 — Synthetic Kick & Blowout Scenario Dataset (Sample)" size_categories: - 100K pit_volume += influx_bbl × min(1.0, progress × 1.55) > flow_out += 8 + influx_bbl × min(1.0, progress × 2.2) > spp -= 40 + influx_bbl × progress × 1.5 > gas_units += 30 + influx_bbl × progress × 6 > mud_weight -= min(1.6, influx_bbl/120) × progress This produces **classic kick signatures detectable from the timeseries**: flow_out > flow_in, rising pit volume, falling standpipe pressure, rising gas units. Critical for training kick-detection ML on realistic sensor patterns rather than just labeled-event metadata. **`bop_operations.csv`** — every kick triggers all 5 preventers per the **API RP-53 + IADC BOP stack convention**: | Preventer | Closure Time (s) | Function | |---|---:|---| | Annular | 18 ± 4 | Soft-seal against varied pipe sizes | | Upper pipe ram | 18 ± 4 | Pipe-specific hard seal (upper) | | Lower pipe ram | 18 ± 4 | Pipe-specific hard seal (lower) | | Blind shear ram | 23 ± 4 | Last-resort cut & seal (+5s for shear time) | | Casing shear ram | 18 ± 4 | Cut casing for emergency disconnect | Activation success ~98.5% baseline, dropping to ~95% under escalation ≥4 and ~94% under blowout — matching SINTEF OREDA reliability statistics. **`kick_events.csv`** — 7-class kick taxonomy weighted per Bourgoyne et al.: gas kick 33% / swab kick 15% / loss-kick 12% / HPHT gas 12% / oil 8% / water 10% / shallow gas 10%. Influx volume drawn from lognormal(μ=2.35, σ=0.65) → median ~10 bbl, p90 ~25 bbl, matching industry envelope. **`gas_influx_profiles.csv`** — gas expansion ratio computed from TVD via `1 + tvd/2600` — physics-consistent ideal-gas law approximation (BHP_factor / atm_factor ≈ TVD-dependent scaling). At ~13 kft TVD, expected expansion is ~6× — exactly what the sample produces. **`scenario_labels.csv`** — 4-class risk taxonomy mapped to training class: | Label | Population | Trigger | |---|---:|---| | `normal_drilling` | ~86% | No kick | | `kick` (medium risk) | ~8% | Kick detected, escalation < 4 | | `kick` (high risk) | ~1% | Kick detected, escalation ≥ 4, no blowout | | `blowout` (critical) | ~6% | Catastrophic escalation | The **6% catastrophic rate is intentionally elevated** for safety-training ML (real-world per-well rates are ~0.1-0.3%); see "Honest disclosure" section for the rationale. --- ## Suggested use cases 1. **Kick detection from timeseries** — train binary or sequence classifiers on the 108,000-row drilling_timeseries to predict `kick_detected_flag` from pit / flow / SPP / gas patterns. The post-kick perturbations are physics-coupled, so models will learn real well-control signatures. 2. **BOP reliability ML** — predict `activation_status` and `seal_integrity_score` per preventer from kick severity, blowout risk, and equipment-failure features. 3. **Blowout escalation prediction** — binary classifier on `blowout_escalated_flag` from kick characteristics (influx volume, detection delay, kick type, escalation level). 4. **Kick type classification** — multi-class classifier (7 classes: gas/oil/water/swab/loss-kick/shallow gas/HPHT gas) from pre-detection telemetry features. 5. **Alarm acknowledgment-delay regression** — predict `acknowledgment_delay_sec` from alarm severity, crew training level, and operational context. 6. **Root cause analysis classification** — 10-class classifier on `root_cause_category` from incident features for incident-review automation. 7. **Kill method selection** — predict optimal `kill_method` (6 classes) from well geometry, kick characteristics, and BOP status. 8. **Risk level scoring** — 4-class ordinal classifier on `risk_level` (low/medium/high/critical) from upstream features for real-time well-control dashboards. 9. **Multi-table relational ML** — entity-resolution and graph-based learning across the 13 joinable tables via `well_id` and `kick_id`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil011-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/oil011-sample/wells_master.csv") ts = pd.read_csv("hf://datasets/xpertsystems/oil011-sample/drilling_timeseries.csv") kicks = pd.read_csv("hf://datasets/xpertsystems/oil011-sample/kick_events.csv") bop = pd.read_csv("hf://datasets/xpertsystems/oil011-sample/bop_operations.csv") labels = pd.read_csv("hf://datasets/xpertsystems/oil011-sample/scenario_labels.csv") # Join timeseries to kick events ts_with_kicks = ts.merge(kicks, on="kick_id", how="left") ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving both `random.seed` and `np.random.default_rng`). 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 for ML prototyping and safety-training research. A few important notes: 1. **Elevated kick and blowout rates are intentional for safety-training data.** The sample's per-well kick rate (~14%) and blowout-given-kick rate (~43%, vs IOGP global ~3-7%) are deliberately amplified via the `blowout_escalation_rate=0.38` parameter. This produces a **balanced training set with sufficient positive examples** of all escalation levels — useful for ML, but **not representative of true field-rate well-control incident frequencies**. For epidemiological modeling of well-control incidents, scale down events to match IOGP Report 432 baseline rates. 2. **Successful-kill rate is ~53%, lower than industry success rates (~85-95%).** This is a consequence of (1) above: with ~43% of kicks escalating to blowout in the generator, `kill_success` is constructed as `not blowout AND U(0,1) < 0.93`. For ML training on kill-method selection this is fine; for kill-success forecasting, re-weight the training set. 3. **BOP activation timing is fast end of the API envelope.** Mean closure time is ~18s, matching modern subsea BOPs but at the aggressive end of the API 16D spec (annular <30s, ram <45s). Older land-rig BOP stacks may run 25-40s. Adjust calibration if your target audience is legacy onshore equipment. 4. **Each kick triggers all 5 preventers** — generates 5 BOP rows per kick deterministically. Real BOP procedures activate based on escalation state (annular first, then rams, shear rams last). The sample provides activation timing for every preventer regardless of whether it would be procedurally invoked, useful for individual- preventer reliability ML but not for activation-sequence modeling. 5. **Time-series anomaly flag is uniform 3.1% Bernoulli** across all timepoints (per `anomaly_injection_rate`) — does not concentrate around kick events. For anomaly-detection ML, the `kick_id` non-NONE timepoints are the true positive signal channel; the `anomaly_flag` is sensor noise / data-quality drill. 6. **`kick_severity` distribution is bimodal** (peaks at 1-2 and 4-5) because blowout-escalated kicks override severity to {4, 5}. This reflects the generator's coupling between blowout-flag and severity, not a modeling error. For severity regression, condition on `blowout_escalated_flag` to avoid label leakage. 7. **Equipment failures are sparse** (~12% of kicks have a failure recorded), reflecting the conditional injection rate `equipment_failure_rate=0.019 + 0.25*blowout + 0.05*high_escalation`. For failure-mode ML, expect class-imbalanced learning conditions. --- ## Full product The **full OIL-011 dataset** ships at **50,000 wells × 15 timepoints** (prod mode) with field-rate-calibrated incident frequencies (IOGP-matched 0.3-0.5% blowout rate), legacy-equipment BOP timing distributions, and procedurally-correct BOP activation sequencing — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil011_sample_2026, title = {OIL-011: Synthetic Kick & Blowout Scenario Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil011-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-21 23:49:57 UTC - Wells : 1,500 - Timepoints/well : 72 - Basins : 12 (Permian, Gulf of Mexico, North Sea, Brazil Pre-Salt, Middle East, Eagle Ford, Bakken, West Africa, North Sea HPHT, Arctic, North African, Haynesville) - Formations : 10 (overpressured shale, fractured carbonate, deepwater turbidite, HPHT gas sand, sour gas carbonate, etc.) - Kick types : 7 (gas, oil, water, swab, loss-kick, shallow gas, HPHT gas) - Blowout types : 5 (surface, underground, riser gas, wellhead, ignited) - BOP preventers : 5 (annular, upper pipe ram, lower pipe ram, blind shear ram, casing shear ram) - Kill methods : 6 (driller, wait-and-weight, bullheading, volumetric, lubricate-and-bleed, dynamic kill) - Calibration basis : API 16D, API RP-53, API RP-59, IADC WellSharp, IADC Well Control taxonomy, IOGP Report 432, NORSOK D-010, SINTEF OREDA, Bourgoyne et al. (1986) - Overall validation: 100.0/100 — Grade A+