--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - midstream - flow-assurance - pipeline-integrity - wax-deposition - hydrate-formation - asphaltene-precipitation - multiphase-flow - thermal-management - xpertsystems pretty_name: "OIL-015 — Synthetic Flow Assurance Dataset (Sample)" size_categories: - 100K= WAT (physics: no wax deposition above WAT). Validates the generator's threshold-gating logic. | | 6 | asphaltene threshold gating fidelity | 1.0000 | 0.99 | ±0.02 | ✓ PASS | SPE 28994 (Hammami & Raines) asphaltene precipitation thermodynamics — fraction of rows where precipitation_rate is correctly zero when P_fluid >= AOP (physics: asphaltenes stay in solution above AOP). Validates the generator's threshold gating. | | 7 | wax deposition pearson correlation | 0.7894 | 0.7 | ±0.2 | ✓ PASS | Pedersen et al. (1991) + Hammami & Raines (1997) — expected positive correlation between (WAT − T_fluid) and deposition rate when delta-T > 0 (physics: greater subcooling drives faster crystallization). Validates wax deposition rate scales with thermodynamic driving force. | | 8 | asphaltene precipitation pearson correlation | 0.7534 | 0.65 | ±0.2 | ✓ PASS | SPE 28994 (Hammami & Raines, 1997) — expected positive correlation between (AOP − P_fluid) and precipitation rate when delta-P > 0 (physics: greater pressure deficit drives faster asphaltene flocculation). Validates precipitation rate scales with thermodynamic driving force. | | 9 | flow regime diversity entropy | 0.9999 | 0.99 | ±0.03 | ✓ PASS | Mandhane et al. (1974) + Beggs & Brill (1973) multiphase flow regime classification — 5-class flow regime diversity benchmark (slug, annular, bubble, stratified, churn), normalized Shannon entropy. ML training portfolios typically use uniform sampling across regimes. | | 10 | pipeline type diversity entropy | 0.9959 | 0.97 | ±0.04 | ✓ PASS | Rystad Energy + IHS Markit global pipeline tracker — 6-class pipeline-type diversity benchmark (deepwater subsea, heavy oil gathering, gas condensate, LNG feed, offshore export, shale multiphase), normalized Shannon entropy. | **Overall: 100.0/100 — Grade A+** (10 PASS · 0 MARGINAL · 0 FAIL of 10 metrics) --- ## Schema highlights **`wax_deposition.csv`** — implements **NACE TM0274** wax-thermodynamic threshold gating: > wat = N(118, 8) °F > deposition_rate = max(0, WAT − T_fluid) × U(0.01, 0.15) mm/day This means **wax deposition is exactly zero when T_fluid ≥ WAT** (per Pedersen et al. 1991 thermodynamics — wax stays in solution above WAT) and scales positively with subcooling below WAT. The sample observes ~10% of rows in the deposition zone, matching the realistic fraction of pipeline operations that drop below WAT. **`asphaltene_precipitation.csv`** — implements **SPE 28994 (Hammami & Raines)** asphaltene-thermodynamic threshold gating: > aop = N(3500, 600) psi > precipitation_rate = max(0, AOP − P_fluid) × U(0.0001, 0.005) Asphaltenes precipitate **only when P_fluid < AOP** (per asphaltene solubility thermodynamics — asphaltenes stay in solution above onset pressure). The sample observes ~18% of rows in the precipitation zone, matching the realistic fraction of pipeline operations that drop below AOP. **`thermal_profiles.csv`** — heat loss conditioned on delta-T per **ISO 13703** subsea pipeline thermal design: > seabed_temp = N(40, 10) °F > heat_loss = max(0, T_fluid − T_seabed) × U(1, 10) BTU Heat loss only occurs when fluid is warmer than seabed (always true in this sample given fluid temp ~145°F vs seabed ~40°F). **`multiphase_flow.csv`** — 5-class flow regime classification per **Mandhane et al. (1974) / Beggs & Brill (1973)** flow-regime maps: slug / annular / bubble / stratified / churn. Sample distribution is near-uniform (~20% each) for ML-balanced classification training. **`integrity_risk_labels.csv`** — 3-class integrity grade derived from blockage probability: | Grade | Trigger | |---|---| | `LOW` | risk ≤ 0.45 | | `MEDIUM` | 0.45 < risk ≤ 0.75 | | `HIGH` | risk > 0.75 | --- ## Suggested use cases 1. **Wax deposition rate regression** — predict `deposition_rate_mm_day` from operating conditions (temperature/pressure/flow_rate) and WAT. **Strong physics signal**: threshold gating + delta-T correlation r ≈ 0.79. 2. **Asphaltene precipitation regression** — predict `precipitation_rate` from pressure conditions and AOP. **Strong physics signal**: threshold gating + delta-P correlation r ≈ 0.75. 3. **Flow regime classification** — multi-class (5-way) classifier on `flow_regime` from gas/liquid fractions + slug frequency features. 4. **Hydrate inhibition state classification** — 3-class (active/ partial/failed) classifier for inhibition effectiveness ML. 5. **Pipeline integrity grading** — 3-class ordinal classifier on `integrity_grade` (LOW/MEDIUM/HIGH) — useful as label-only reference; see Honest Disclosure §3 for feature-engineering caveats. 6. **Chemical injection optimization** — regression on `effectiveness_pct` from inhibitor type + dosage features for chemical program tuning. 7. **Heat loss prediction** — regression on `heat_loss_btu` from fluid/seabed temperature + pipeline characteristics. Anchors to ISO 13703 thermal design. 8. **Multi-table relational ML** — entity-resolution and graph neural-network learning across the 9 joinable tables via `pipeline_id` + `condition_id`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil015-sample", data_files="wax_deposition.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd pipelines = pd.read_csv("hf://datasets/xpertsystems/oil015-sample/pipelines_master.csv") conditions = pd.read_csv("hf://datasets/xpertsystems/oil015-sample/operating_conditions.csv") wax = pd.read_csv("hf://datasets/xpertsystems/oil015-sample/wax_deposition.csv") asp = pd.read_csv("hf://datasets/xpertsystems/oil015-sample/asphaltene_precipitation.csv") # Wax deposition is keyed by condition_id (embedded in wax_id): wax["condition_id"] = wax["wax_id"].str.replace("WAX-", "", regex=False) wax_joined = wax.merge(conditions, on="condition_id") # Now you have WAT + T_fluid + P_fluid features ready for ML ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving both `random.seed` and `np.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 for flow assurance ML research, not for live pipeline operations decisions. Several important notes: 1. **Operating conditions are independent of pipeline characteristics.** The generator samples pressure/temperature/flow_rate from fixed Gaussians, *not* conditioned on pipeline diameter, water depth, design pressure, or insulation type. This is a deliberate simplification for ML-balanced training but is not realistic — real deepwater pipelines run colder than onshore (better heat transfer to seabed), and design pressure limits operating pressure. For physics-realistic conditioning ML, treat operating conditions as features and pipeline characteristics as separate features rather than expecting cross-table coupling. 2. **Hydrate events have no thermodynamic gating.** Real hydrate formation occurs **inside the hydrate envelope** (low T, high P region per Sloan & Koh 2008). The generator samples `hydrate_risk_score` from `N(0.35, 0.15)` independent of T/P, so hydrate risk is not physically coupled to operating conditions. **This is a significant simplification.** For hydrate ML that requires T-P-conditioned risk, use a Sloan & Koh CSMHYD-style envelope calculation on top of operating conditions, or wait for the full product v1.1 which will introduce envelope-aware hydrate gating. 3. **`integrity_risk_labels.csv` is feature-decoupled.** The blockage probability is sampled from `N(0.25, 0.15)` independent of wax deposition rate, asphaltene precipitation, hydrate risk, or other upstream features. **Models trained to predict `integrity_grade` from upstream features will not learn meaningful patterns** because the label is not a function of the features. This is a generator design choice (likely placeholder for future coupling). For integrity-grade ML, build a derived label from weighted combinations of `wax_thickness_mm`, `hydrate_risk_score`, `precipitation_rate`, and `heat_loss_btu` rather than using the sampled label directly. The integrity_risk_labels table is best used as a **reference distribution** for production label calibration, not as a supervised ML target. 4. **Chemical injection effectiveness is uniform-sampled.** The `effectiveness_pct` column from `U(40, 99)` is not tied to inhibitor type, dosage, or operating conditions — so an "MEG @ 100 ppm in a wax-deposition scenario" might show 95% effectiveness in the sample, which is physically wrong (MEG is a hydrate inhibitor, not a wax inhibitor). **Inhibitor-effectiveness ML on this sample will learn marginals, not couplings.** For physically-correct inhibitor-effectiveness ML, post-process the data to align inhibitor type with deposition type, or wait for the full product v1.1. 5. **Wax/asphaltene rates are dimensionless multipliers**, not physically calibrated. The wax `U(0.01, 0.15)` multiplier and asphaltene `U(0.0001, 0.005)` multiplier produce rates in the correct order of magnitude (mm/day for wax, dimensionless flux for asphaltene) but are not calibrated to specific crude compositions. For absolute-rate prediction, the labels need recalibration against the user's crude assay; for relative ranking ML (e.g., "which pipeline is most at risk"), the relative ordering is preserved. 6. **No time-series autocorrelation across hourly steps.** Each row in `operating_conditions.csv` is sampled independently — there's no Markov / AR / drift modeling across consecutive hours. **Time- series ML that relies on temporal smoothness will not get realistic signal from this data.** Treat the sample as a panel of independent observations, not as time-series. 7. **Pipeline length / diameter are uniformly distributed**, not conditioned on pipeline type. Real LNG feed lines run 36"+ for high-volume gas transport; real heavy-oil gathering lines are typically 6-12". The sample uses U(4, 36) inches across all types. For type-conditional ML, post-process the data with industry-standard pipeline-sizing priors. --- ## Full product The **full OIL-015 dataset** (in development) will ship at 5,000+ pipelines × 8,760 hourly records (1 full year) with **physics-conditioned hydrate envelope gating** (Sloan & Koh CSMHYD-style), **type-conditional pipeline sizing**, **coupled integrity-risk labels** derived from upstream deposition features, and **inhibitor-specific effectiveness coupling** — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil015_sample_2026, title = {OIL-015: Synthetic Flow Assurance Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil015-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-22 13:17:09 UTC - Pipelines : 300 - Rows per pipeline : 100 (~hourly for ~4 days) - Pipeline types : 6 (deepwater subsea, heavy oil gathering, gas condensate, LNG feed, offshore export, shale multiphase) - Insulation types : 5 (wet insulation, pipe-in-pipe, foam, none, vacuum insulated) - Flow regimes : 5 (slug, annular, bubble, stratified, churn) - Inhibitor types : 4 (MEG, methanol, wax inhibitor, asphaltene dispersant) - Calibration basis : NACE TM0274, NACE SP0775, Sloan & Koh (2008), SPE 28994 (Hammami & Raines), Mandhane (1974), Beggs & Brill (1973), API RP-14E, ISO 13703, DNV-RP-F101, Pedersen (1991), Rystad, IHS - Overall validation: 100.0/100 — Grade A+