--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - artificial-lift - esp - gas-lift - rod-pump - predictive-maintenance - failure-prediction - nodal-analysis - xpertsystems pretty_name: "OIL-014 — Synthetic Artificial Lift Dataset (Sample)" size_categories: - 100K load_factor = (frequency_hz / 60) × (liquid_rate / 1100) > efficiency = 76 − 12·|load_factor − 1.0| − 8·gas_lock − 4·scale − 3·sand + noise > gas_lock_probability = sigmoid((GOR − 1800) / 550) × 0.65 Voltage centered on **2400 V** (3-phase ESP motor class per ANSI C50.41 + IEEE 141), frequency 35-72 Hz (modern VFD operational envelope). The ESP frequency↔load Pearson correlation is r ≈ 0.94 in the sample — strong physics coupling validates the design. **`rod_pump_operations.csv`** — beam pump physics per API 11L: > SPM = 5.5 + 0.005·liquid_rate − 0.014·water_cut + noise > fillage = 88 − 0.018·GOR − 12·sand − 0.09·water_cut + noise > fluid_pound_prob = sigmoid((62 − fillage) / 7) + noise > polished_rod_load = 3500 + 1.7·depth + 6.5·liquid + noise Fillage↔fluid pound Pearson correlation is r ≈ −0.97 — **near-perfect inverse coupling** per API RP-11ER physics (fluid pound is caused by incomplete pump fillage). **`gas_lift_operations.csv`** — gas lift physics per SPE 14253: > casing_pressure = 350 + 1.05·injection_rate + noise > tubing_pressure = 120 + 0.55·casing_pressure − 0.04·liquid + noise > valve_depth = depth × U(0.42, 0.88) > active_valves = round(depth / 2500 + noise) > lift_gas_utilization = 0.86 − 0.14·|load_factor − 0.85| − 0.05·scale + noise Active valve count averages ~4 — matches API RP-19G gas-lift completions for moderately-deep (8-12 kft) wells. **`nodal_analysis.csv`** — IPR-VFP intersection point per **SPE 174021** nodal analysis convention. Sample mean nodal balance error is ~1.85% — **well below the 5% production-grade target** for IPR-VFP equilibrium models. **`power_consumption.csv`** — power-factor-corrected 3-phase electrical metrics: > kW = lift_specific_base + lift_specific_coefficient · liquid_rate + noise > amperage = kW × 1000 / (√3 × voltage × power_factor) Voltage is **lift-type stratified per ANSI C50.41**: ESP 2400V (HV motor class), Gas Lift 480V (LV instrumentation/controls), Rod Pump 480V (LV prime mover). --- ## Suggested use cases 1. **ESP failure prediction** — binary classifier on `failure_risk_score > threshold` from ESP operations features (frequency / load factor / intake pressure / efficiency / gas lock probability). 2. **Rod pump fluid-pound detection** — binary classifier on `fluid_pound_probability > 0.5` from fillage / sand / water cut features. Strong physics signal (r ≈ −0.97 fillage↔fluid pound). 3. **Lift type recommendation** — multi-class (3-way: ESP/Gas Lift/ Rod Pump) classifier from well characteristics (depth / reservoir P / GOR / water cut / well class). 4. **Pump-off detection** — binary classifier on `pump_off_probability` for rod pump wells, from SPM / stroke / fillage features. 5. **Gas lock detection** — binary classifier on `gas_lock_probability > 0.5` from GOR / intake pressure / load factor features. 6. **Optimization-priority ranking** — regression on `recommendation_score` from upstream operations features. 7. **Remaining-useful-life regression** — predict `remaining_useful_life_days` from health + operations + well metadata features (standard PHM/RUL benchmark). 8. **3-class risk-grade classification** — ordinal classifier on `failure_risk_grade` (low/medium/high) from upstream features. 9. **Nodal-analysis fitting** — regression on `operating_point_bpd` from IPR slope + drawdown + VFP pressure loss features. Anchors to SPE 174021 production-grade <5% nodal error. 10. **Multi-table relational ML** — entity-resolution and graph neural-network learning across the 12 joinable tables via `well_id` + `timestamp`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil014-sample", data_files="esp_operations.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd wells = pd.read_csv("hf://datasets/xpertsystems/oil014-sample/wells_master.csv") esp = pd.read_csv("hf://datasets/xpertsystems/oil014-sample/esp_operations.csv") rp = pd.read_csv("hf://datasets/xpertsystems/oil014-sample/rod_pump_operations.csv") gl = pd.read_csv("hf://datasets/xpertsystems/oil014-sample/gas_lift_operations.csv") labels = pd.read_csv("hf://datasets/xpertsystems/oil014-sample/artificial_lift_labels.csv") # Filter wells by lift type and join to operations esp_wells = wells[wells["lift_type"] == "ESP"] esp_joined = esp.merge(esp_wells, on="well_id") ``` --- ## 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 calibrated for artificial-lift ML research, not for live lift optimization decisions. A few notes: 1. **Risk-grade distribution skews toward "low" and "medium"** — the sample has ~77% low / ~23% medium / 0% high risk-grade labels. The "high" threshold (risk ≥ 0.70) requires a confluence of high corrosion + high scale + high sand + old install age, which is rare in the Beta-distributed sample at this scale. The full product (120K wells × 90 days) generates enough tail wells to populate the "high" class. For ML at sample scale, use `failure_risk_score` as a continuous regression target rather than the 3-class label. 2. **Failure events are rare** (~6% of wells produce a failure event in the 60-day window). This matches real-world artificial-lift reliability (annual ESP failure rates ~10-20%, gas lift ~5-10%, rod pump ~15-25% per SPE 174021 / Rystad data). For class- balanced ML training, oversample positive cases or use the continuous `failure_risk_score` columns in `esp_operations`, `rod_pump_operations`, `equipment_health`, and `artificial_lift_ labels`. 3. **Two distinct failure-risk-score scales coexist.** Time-series tables (esp / rod_pump / equipment_health) use a sigmoid model centered around ~0.03 mean (per-timestep instantaneous risk). The labels table uses an additive index averaging ~0.34 mean (well-aggregated lifetime risk). These are **not the same metric** — they're computed differently and serve different ML purposes (instantaneous detection vs lifetime prognosis). Don't mix them in a single regression target. 4. **All voltage in `power_consumption.csv` is lift-type stratified only by class**, not by well-specific motor sizing. Real ESP installations use 2300V / 3300V / 4160V depending on motor HP and depth — the sample uses N(2400, 180) for all ESP wells. For motor-sizing ML, condition the voltage feature on depth and liquid rate. 5. **The optimization_recommendations table is uniformly sampled**, not tied to specific underlying conditions. Real production engineering recommends specific interventions based on observed symptoms (high gas lock → reduce frequency, high fluid pound → reduce SPM). For optimization ML, treat this table as label- only and engineer features from the operations tables. 6. **Production decline is linear-on-time** (`depletion = 1 - 0.22 × t/n_steps`), not Arps-driven. For decline-curve ML, use OIL-013 (which implements full Arps hyperbolic decline). OIL-014 focuses on lift-system optimization and instantaneous performance, not long-horizon decline. 7. **Sample-scale ESP class is over-represented vs the global declared 42/33/25 mix.** Sample observed 44/21/35 (ESP/Gas Lift/ Rod Pump). The basin-class conditioning produces this because the sample's basin draws favor unconventional/tight/heavy oil classes (which favor rod pump and ESP), giving fewer Gas Lift wells than declared. The full product (120K wells) gives the declared 42/33/25 split with basin-conditioning preserved. --- ## Full product The **full OIL-014 dataset** ships at **120,000 wells × 90 days** (prod mode) producing several hundred million operation records with substantial failure / maintenance / optimization event populations, full "high" risk- grade class population, and per-motor-sized voltage stratification — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil014_sample_2026, title = {OIL-014: Synthetic Artificial Lift Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil014-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-22 12:49:51 UTC - Wells : 400 - Days simulated : 60 - Frequency : 24h (60 timesteps per well) - Basins : 10 (Permian, Eagle Ford, Bakken, GoM, North Sea, Middle East, Canadian Heavy Oil, Williston, Anadarko, San Joaquin) - Well classes : 5 (unconventional oil, tight oil, offshore, carbonate, heavy oil) - Lift types : 3 (ESP, Gas Lift, Rod Pump) basin-class-conditioned - Failure modes : 19 (ESP: 7, Gas Lift: 6, Rod Pump: 6) - Optimization types: 9 (lift-conditioned recommendations) - Calibration basis : SPE 174021, API RP-11ER, API 11L, API 11AX, SPE 14253, API 670, IEEE 141, ANSI C50.41, IEC 60038, Rystad, Spears, Schlumberger Reda, Centrilift, Lufkin design literature - Overall validation: 100.0/100 — Grade A+