--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - multiphase-flow - beggs-brill - flow-regime-classification - slugging-prediction - pvt-properties - separator-performance - xpertsystems pretty_name: "OIL-018 — Synthetic Multi-Phase Flow Dataset (Sample)" size_categories: - 100K BHP states. | | 5 | avg separator oil recovery pct | 96.0900 | 95.0 | ±3.0 | ✓ PASS | API 12J (Specification for Oil and Gas Separators) + GPSA Engineering Data Book — typical oil recovery for production separators (93-98% for properly-sized vessels with good retention time) | | 6 | avg separator gas efficiency pct | 93.4835 | 93.0 | ±3.0 | ✓ PASS | API 12J + GPSA Engineering Data Book — typical gas separation efficiency (90-96% for vertical/horizontal separators with mist extractor) | | 7 | avg gas z factor | 0.8609 | 0.86 | ±0.08 | ✓ PASS | Standing-Katz (1942) gas compressibility chart — typical Z-factor for natural gas at reservoir P/T conditions (0.75-0.95 for most production scenarios; 0.86 is the median for moderate-pressure portfolios) | | 8 | hydrate prone risk coupling | 0.1554 | 0.1 | ±0.05 | ✓ PASS | Sloan & Koh (2008) 'Clathrate Hydrates of Natural Gases' — expected positive difference in hydrate risk score between hydrate-prone wells and non-hydrate-prone wells (validates flag-conditioned risk physics; generator coefficient is 0.22 prone-flag boost) | | 9 | wax prone risk coupling | 0.1796 | 0.15 | ±0.05 | ✓ PASS | NACE TM0274 (Wax Appearance Temperature Measurement) + Pedersen et al. (1991) — expected positive difference in wax risk score between wax-prone wells and non-wax-prone wells (validates flag-conditioned risk physics; generator coefficient is 0.18 prone-flag boost) | | 10 | basin diversity entropy | 0.9812 | 0.95 | ±0.05 | ✓ PASS | Rystad Energy + IHS Markit + EIA global production tracker — 12-class basin diversity benchmark (Permian Delaware/Midland, Eagle Ford, Bakken, Marcellus, GoM Deepwater, North Sea, Brazil Pre-Salt, Middle East Carbonate, West Africa, Canadian Heavy Oil, North African Carbonate), normalized Shannon entropy | **Overall: 100.0/100 — Grade A+** (10 PASS · 0 MARGINAL · 0 FAIL of 10 metrics) --- ## Schema highlights **`03_multiphase_flow_timeseries.csv`** — the production spine with **Hagedorn & Brown (1965) vertical multiphase flow physics**: > WHP = P_res − 0.42·depth − 0.00095·liquid_bpd − 0.018·gas_mscfd + noise > BHP = WHP + 0.38·depth + 0.0006·liquid_bpd > vsl (ft/s) = liquid_bpd × 5.615 / 86400 / area_ft2 > vsg (ft/s) = gas_mscfd × 1000 / 86400 / area_ft2 × (520/T_R) × (14.7/WHP_psia) The vsg formula applies **gas density correction at standard conditions** (14.7 psia / 520 R = 60°F) per **GPSA Engineering Data Book** convention. **BHP > WHP is enforced for 100% of timesteps** — physical consistency validates the pressure model. **`05_flow_regimes.csv`** — **Beggs & Brill (1973) inclined pipe regime classification**: | Regime | Trigger | |---|---| | `bubble` | vsg/vsl < 0.18 AND vsl ≥ 0.20 | | `slug` | 0.18 ≤ vsg/vsl < 0.38 AND vsl ≥ 0.30 | | `churn` | 0.38 ≤ vsg/vsl < 0.70 | | `annular` | vsg/vsl ≥ 0.70 AND vsl ≥ 0.25 | | `stratified` | vsl < 0.75 AND vsg < 0.75 AND \|inclination\| < 6° | | `mist` | vsg/vsl ≥ 0.85 AND vsl < 0.50 | | `transition` | (fallback) | **`08_pvt_properties.csv`** — **Vasquez & Beggs (1980)** PVT correlations: > bubble_point = 2.2 × GOR + noise (Lasater 1958 form) > oil_viscosity = 14.5 / API × (1 + 0.00035 × max(P_bp − P, 0)) > Z_factor = N(0.86, 0.06) clip(0.62, 1.08) (Standing-Katz envelope) > oil_FVF = 1 + GOR/6500 + noise (Vasquez-Beggs form) **`07_separator_performance.csv`** — **API 12J** separator design metrics: > oil_recovery = 0.965 − 0.0012·max(water_cut − 40, 0) + noise > gas_efficiency = 0.955 − 0.000045·sep_pressure + noise Sample mean oil recovery 96.1%, gas efficiency 93.5% — **within API 12J production-grade specifications** (93-98% oil, 90-96% gas). **`09_hydrate_wax_risk.csv`** — **Sloan & Koh (2008)** hydrate physics + **NACE TM0274** wax thermodynamics with **flag-conditioned risk amplification**: > hydrate_risk = 0.18 + 8e-5·P − 4e-3·(T − 60) + 0.22·hydrate_prone_flag + noise > wax_risk = 0.12 + 0.45·(T < WAT) + 0.18·wax_prone_flag + noise Hydrate risk **increases with pressure and decreases with temperature** per hydrate stability zone physics. Wax risk uses a hard threshold below WAT (Wax Appearance Temperature). Both have **flag-conditioned amplification** that ML models should learn. **`12_production_labels.csv`** — 4-class stability grade per multiphase flow operability convention: | Grade | Stability score | |---|---| | `A` | ≥ 0.80 | | `B` | 0.60 ≤ score < 0.80 | | `C` | 0.35 ≤ score < 0.60 | | `D` | < 0.35 | --- ## Suggested use cases 1. **Flow regime multiclass classification** — predict `flow_regime` (5-7 classes) from vsl/vsg/inclination features. **Strong physics signal**: classification follows **Beggs & Brill (1973)** deterministic regime boundaries. 2. **Slugging detection** — binary classifier on `slugging_flag` from regime + holdup + pressure features. 3. **Liquid loading prediction** — binary classifier on `liquid_loading_flag` per **Turner et al. (1969)** criterion (vsg < 2.5 ft/s AND water_cut > 0.45). 4. **PVT property regression** — predict bubble point / FVF / Z-factor from upstream features (depth, API gravity, GOR). Strong physical signal per Vasquez-Beggs (1980). 5. **Separator efficiency optimization** — regression on `oil_recovery_pct` and `gas_efficiency_pct` from upstream feed conditions. Anchors to API 12J production-grade targets. 6. **Hydrate / wax risk scoring** — regression on `hydrate_risk_score` / `wax_risk_score` from operating P/T + integrity flags. **Strong flag-conditioned coupling**: prone-flag risk amplification is 0.15-0.22 in the sample. 7. **10-class anomaly type classification** — multi-class classifier on `anomaly_type` from upstream features. 8. **4-class production stability classification** — ordinal classifier on `production_stability_grade` (A/B/C/D); see Honest Disclosure §3 for the class-imbalance caveat. 9. **Pressure gradient prediction** — regression on `pressure_gradient_psi_per_ft` from depth + fluid composition features per Hagedorn & Brown (1965). 10. **Multi-table relational ML** — entity-resolution and graph neural-network learning across the 12 joinable tables via `well_id` + `pipeline_id` + `timestamp`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil018-sample", data_files="03_multiphase_flow_timeseries.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd wells = pd.read_csv("hf://datasets/xpertsystems/oil018-sample/01_wells_master.csv") ts = pd.read_csv("hf://datasets/xpertsystems/oil018-sample/03_multiphase_flow_timeseries.csv") regimes = pd.read_csv("hf://datasets/xpertsystems/oil018-sample/05_flow_regimes.csv") labels = pd.read_csv("hf://datasets/xpertsystems/oil018-sample/12_production_labels.csv") # Join timeseries with well metadata for ML feature engineering joined = ts.merge(wells, on="well_id") # Flow regime classification training set: X = regimes[["superficial_velocity_liquid_ft_s", "superficial_velocity_gas_ft_s", "inclination_deg"]] y = regimes["flow_regime"] ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving `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 multiphase flow / production engineering ML research, not for live operational decisions. Several important notes: 1. **Flow regime is heavily slug-dominant (~57%) and churn-dominant (~26%).** The Beggs & Brill (1973) regime classifier uses superficial velocity ratios that naturally produce lots of slug+churn at typical wellhead conditions (moderate vsl, moderate vsg). The `slugging_share_estimate` in the generator's metrics.json reports 0.67 — physically correct for moderate-rate wells but means **annular and mist regimes are underrepresented (~0.1% and ~0% respectively)**. For class-balanced flow regime ML, oversample annular/mist or filter for high-velocity wells. 2. **Liquid loading flag fires at ~38%** because **Turner et al. (1969)** criterion `vsg < 2.5 ft/s AND water_cut > 0.45` triggers for many mature water-cut wells. This is **realistic for late-life production** but means liquid loading ML on this sample is biased toward mature wells. For greenfield liquid-loading ML, filter the timeseries to early-life timesteps (`day < 30`). 3. **Production stability grade is ~93% A** because the stability formula `1 − 0.55·slug_flag − 0.25·(water_cut > 0.75)` rarely drops below 0.80 in 60-day simulations. **4-class stability classification is heavily imbalanced at sample scale.** Use `optimization_candidate_score` as a continuous regression target instead. 4. **Reservoir pressure ranges from 1500 to 18000 psi** in the wells table because the generator adds `0.42·depth_ft` to the base, so deep wells (28000 ft cap) produce 13000+ psi reservoir P. **The `mean_pressure_psi: 4200` config parameter is the *intercept*, not the actual mean.** Observed mean is ~8450 psi due to the depth amplification — this is realistic for a portfolio with deep deepwater wells. 5. **Oil rate compound-lognormal-skew effects.** The `lognormal(log(1800), 0.65)` baseline + decline + season + noise produces observed mean ~2050 bopd vs declared 1800 — a 14% positive bias from compound noise sources. Realistic for production distributions (positive skew) but disclosed. 6. **The `04_pressure_temperature_profiles.csv` table is a snapshot**, not a timeseries — one profile per well at fixed depth points. For pressure-traverse-over-time ML, use `03_multiphase_flow_timeseries.csv` WHP/BHP columns instead. 7. **PVT properties are sampled per well, not per pressure step.** Each well gets 8 PVT samples at randomized pressures (25-115% of reservoir P), not a full PVT envelope. For full bubble-point-vs-pressure curve modeling, use the field-development simulation tools downstream of this dataset. 8. **Anomaly types are uniformly sampled** (~10% each) across 10 classes, not feature-conditioned. Real anomaly distributions are heavily skewed (sensor drift dominates, severe slugging rarer). Treat anomaly_type as label-only at sample scale; full product will add feature-conditioned anomaly priors. --- ## Cross-references to other XpertSystems OIL SKUs This SKU specializes in **multiphase flow / pipeline dynamics**. Related SKUs cover complementary aspects: | SKU | Focus | Use Case | |---|---|---| | **OIL-013** | Production engineering | Daily production with downtime/anomaly events at single-well scale | | **OIL-014** | Artificial lift performance | ESP / Gas Lift / Rod Pump operations per-period | | **OIL-015** | Flow assurance | Pipeline-only wax/hydrate/asphaltene threshold-gated deposition (midstream) | | **OIL-018** | Multi-phase flow | Beggs-Brill regime classification + PVT + separator + per-well timeseries (this SKU) | **OIL-018 vs OIL-015**: OIL-015 is **midstream pipeline-only flow assurance** (wax / hydrate / asphaltene threshold gating). OIL-018 is **upstream wellbore + facility multiphase flow** (regime classification, PVT, separator, lift behavior). Use OIL-015 for pipeline integrity ML, OIL-018 for well-to- separator flow modeling ML. --- ## Full product The **full OIL-018 dataset** ships at **40,000 wells × 3,650 days × 15-min resolution** (prod mode) producing several hundred million timeseries rows with **feature-conditioned anomaly priors**, **proper class-balanced stability grades** (mixed simulation durations), **full annular/mist regime representation** (high-velocity well subset), and **per-timestep PVT envelope modeling** — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil018_sample_2026, title = {OIL-018: Synthetic Multi-Phase Flow Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil018-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-22 13:48:46 UTC - Wells : 250 - Days simulated : 60 - Time interval : 240 min (4h) - Pipelines : 35 × 18 segments each - Basins : 12 (Permian Delaware/Midland, Eagle Ford, Bakken, Marcellus, GoM Deepwater, North Sea, Brazil Pre-Salt, Middle East Carbonate, West Africa, Canadian Heavy Oil, North African Carbonate) - Formation types : 7 (carbonate, sandstone, shale, tight sand, turbidite, heavy oil sand, gas condensate) - Lift types : 5 (natural flow, ESP, gas lift, rod pump, plunger) - Flow regimes : 7 (bubble, slug, churn, annular, stratified, mist, transition) per Beggs & Brill (1973) - Anomaly types : 10 (severe slugging, hydrate plugging, wax restriction, separator instability, sensor drift, choke instability, liquid loading, pipeline leak, ESP gas lock, sand erosion) - Calibration basis : Beggs & Brill (1973), Mukherjee & Brill (1985), Hagedorn & Brown (1965), Turner et al. (1969), Standing-Katz (1942), Lasater (1958), Vasquez & Beggs (1980), API 12J, API RP-14E, Sloan & Koh (2008), NACE TM0274, GPSA, Rystad, IHS Markit - Overall validation: 100.0/100 — Grade A+