xpertsystems/oil013-sample

OIL-013 — Synthetic Production Time-Series Dataset (Sample)

SKU: OIL013-SAMPLE · Vertical: Oil & Gas / Upstream Production Engineering License: CC-BY-NC-4.0 (sample) · Schema version: oil013.v1 Sample version: 1.0.0 · Default seed: 42

A free, schema-identical preview of XpertSystems.ai's enterprise production time-series dataset for decline curve ML, artificial-lift optimization, workover-candidate prediction, and water-breakthrough forecasting. The sample covers 250 wells across 10 global basins and 8 asset types, simulated over 365 days, with 116,154 rows including 91,250 daily production records linked across 11 tables.

---

What's in the box

File	Rows	Cols	Description
wells_master.csv	250	24	Well spine: basin, formation, completion, lift type, Arps decline params (qi, di, b)
daily_production.csv	91,250	17	Per-well-per-day oil/gas/water/water-cut/GOR/pressures/uptime/anomaly
reservoir_pressure.csv	6,750	7	Biweekly pressure tests: reservoir P + BHFP + drawdown + test quality
artificial_lift.csv	13,250	9	Weekly lift performance: ESP frequency/pump efficiency/motor temp/vibration/runtime
downtime_events.csv	491	7	8-class downtime (planned/unplanned/facility/weather/flow assurance/lift/integrity/power)
stimulation_events.csv	1	8	Workover/refrac/acidizing/cleanout/lift change with expected/actual uplift
injection_support.csv	40	9	Producer-injector pairings with response-lag correlation scores
production_allocations.csv	625	6	1-4 producing zones per well, Dirichlet-sampled (sums to 100%)
facility_constraints.csv	200	8	Per-field throughput/gas/water handling limits + constraint severity
flow_assurance_events.csv	47	9	6-class flow assurance (scale/paraffin/hydrate/sand/emulsion/corrosion)
production_labels.csv	3,250	9	Monthly ML labels: 6-class forecast + 4 binary flags (workover/water breakthrough/steep decline/lift limited)

Total: 116,154 rows across 11 CSVs, ~14.8 MB on disk.

---

Calibration: industry-anchored, honestly reported

Validation uses a 10-metric scorecard with targets sourced exclusively to named industry standards: Arps (1945) JPT "Analysis of Decline Curves" (canonical hyperbolic decline equation), SPE Petroleum Engineering Handbook Vol V, SPE 152596 (Unconventional Reservoir Decline Curve Analysis), SPE 167242 (Arps b-factor calibration for unconventional wells), SPE 174021 (ESP performance benchmarks), API RP-11ER (sucker rod pumping system design), EIA Annual Energy Outlook, Rystad ShaleWellCube (unconventional well economics), IHS Markit global production tracker, IOGP allocation standards.

Sample run (seed 42, n_wells=250, simulation_days=365):

#	Metric	Observed	Target	Tolerance	Status	Source
1	avg initial oil rate bopd	1232.3387	1100.0	±400.0	✓ PASS	EIA AEO + Rystad ShaleWellCube — mean initial oil rate for mixed US unconventional + international portfolio (Permian/Eagle Ford ~1500 BOPD IP, deepwater ~2000, heavy oil ~300, shale gas ~200 BOPD condensate)
2	avg initial water cut pct	32.7699	34.0	±10.0	✓ PASS	SPE Petroleum Engineering Handbook Vol V + Rystad — mean initial water cut for mixed onshore/offshore production portfolio (greenfield wells typically 5-25%, mature fields 40-70%)
3	avg initial gor scf bbl	1977.7810	1800.0	±600.0	✓ PASS	SPE PEH Vol V + EIA — mean initial gas-oil ratio across mixed oil/condensate/wet-gas portfolio (Permian ~1500, Marcellus 5000+ condensate, Bakken 1200-2500, heavy oil 200-500 scf/bbl)
4	avg nominal decline rate	0.2300	0.23	±0.08	✓ PASS	SPE 152596 (Unconventional Reservoir Decline Curve Analysis) + SPE 167242 — mean first-year nominal annual decline rate for mixed shale/conventional portfolio (shale 0.30-0.65 yr1, conventional 0.05-0.20, deepwater 0.08-0.25)
5	avg arps b factor	0.9574	1.0	±0.3	✓ PASS	Arps (1945) JPT + SPE 167242 — mean hyperbolic exponent b-factor for unconventional/conventional mix (shale typically 1.0-1.8 transitioning to exponential at terminal decline, conventional 0.3-1.0)
6	arps decline fidelity score	0.9501	0.9	±0.06	✓ PASS	Arps (1945) JPT canonical decline equation — fidelity of generated daily production rates to the Arps prediction (computed as 1 - mean absolute relative error on anomaly-free days across 50 sample wells, target ≥0.85 indicates strong Arps physics)
7	production mass balance score	1.0000	0.99	±0.01	✓ PASS	Material balance principle — cumulative production should equal sum of daily rates (verifies generator's cumulative_oil_bbl column is internally consistent, target ≥0.98 indicates proper integration)
8	allocation completeness score	1.0000	1.0	±0.02	✓ PASS	SPE production allocation guidelines + IOGP allocation standards — per-well allocation percentages across producing zones must sum to 100% (validates Dirichlet sampling produces complete allocations)
9	basin diversity entropy	0.9964	0.95	±0.05	✓ PASS	Rystad Energy + EIA + IHS Markit global production tracker — 10-class basin diversity benchmark (Permian, Eagle Ford, Bakken, Marcellus, Haynesville, GoM, North Sea, Middle East, Western Canada, Brazil Pre-Salt), normalized Shannon entropy
10	lift type diversity entropy	0.9078	0.85	±0.1	✓ PASS	API RP-11ER + SPE 174021 + Spears & Associates lift market intelligence — 6-class artificial lift diversity benchmark (natural flow, ESP, rod pump, gas lift, PCP, plunger lift), normalized Shannon entropy (ESP-dominant per industry default weights [0.18, 0.31, 0.22, 0.18, 0.07, 0.04])

Overall: 100.0/100 — Grade A+ (10 PASS · 0 MARGINAL · 0 FAIL of 10 metrics)

---

Schema highlights

daily_production.csv — the production spine, one row per well per day. The oil-rate model is Arps (1945) hyperbolic decline modulated by operational factors:

q(t) = qi / (1 + b·di·t)^(1/b) — Arps hyperbolic oil_rate = q(t) × seasonal × noise × uptime × lift_factor

The decline-curve fidelity is high: at sample scale, the mean absolute relative error between actual rates and pure-Arps predictions on anomaly- free days is ~6%, with the residual driven by lift degradation (built-in) and operational noise (1.5% std). The full Arps physics is preserved well-by-well — see the scorecard for the explicit fidelity check.

reservoir_pressure.csv — biweekly pressure tests with realistic drawdown modeling:

reservoir_pressure(d) = p0 × (1 − 0.22 × d/sim_days) + N(0, 45) bhfp = reservoir_pressure − U(250, 1700) drawdown = reservoir_pressure − bhfp

Pressure-test quality follows the A/B/C grading convention common in production engineering (40% A, 40% B, 20% C).

artificial_lift.csv — weekly performance per lift type. ESP wells get full instrumentation (ESP frequency Hz, pump efficiency, motor temperature F, vibration index); rod pump / PCP / gas lift / plunger get pump efficiency + vibration only. ESP frequency centered at 52 Hz per API/SPE 174021 ESP operating-range guidance.

production_labels.csv — monthly ML labels with 6-class forecast classification:

Class	Trigger
stable	oil_rate ≥ 0.60 × qi AND wc ≤ 62%
moderate_decline	oil_rate < 0.60 × qi
workover_candidate	oil_rate < 0.35 × qi OR wc > 62%
water_breakthrough	wc > 75%
lift_limited	non-natural-flow AND uptime < 78%
steep_decline	oil_rate < 0.20 × qi at early time

Plus four binary flags: workover_candidate_flag, water_breakthrough_flag, steep_decline_flag, lift_limited_flag.

production_allocations.csv — per-well multi-zone allocation using Dirichlet sampling over 1-4 zones; per-well percentages sum to exactly 100%. Allocation methods follow standard production engineering practice: test separator / production logging / model based / commingled estimate (weighted equally).

flow_assurance_events.csv — 6-class flow assurance taxonomy aligned with NACE corrosion standards + SPE flow assurance literature: scale / paraffin / hydrate / sand / emulsion / corrosion. Per-event domain-specific risk indices.

---

Suggested use cases

1. Arps decline curve regression — fit hyperbolic Arps parameters (qi, di, b) from the first 60-180 days of daily production for each well; benchmark against the ground-truth params in wells_master.csv. Strong physics signal — sample mean Arps fidelity is ~94%.
2. 6-class forecast class classification — multi-class classifier on forecast_class from daily production + lift + pressure features.
3. Workover candidate prediction — binary classifier on workover_candidate_flag from upstream features. Highly class-imbalanced (~3% positives), realistic for production engineering operations.
4. Water breakthrough prediction — binary or time-to-event modeling on water_breakthrough_flag from water-cut trajectory features.
5. ESP failure prediction — train RUL or binary failure classifier on artificial_lift.csv ESP-only rows using vibration, motor temperature, pump efficiency degradation as features.
6. Multi-zone allocation regression — predict per-zone allocation percentages from well characteristics and zone metadata.
7. Flow assurance type classification — 6-class classifier on flow_assurance_type from well characteristics and production conditions.
8. Production rate forecasting — N-day-ahead time-series forecasting of oil/gas/water rates from historical features (LSTM / Transformer / TFT benchmark target).
9. Downtime root-cause classification — 6-class classifier on root_cause_category (surface/subsurface/facility/weather/power/ unknown) from production anomaly patterns.
10. Multi-table relational ML — entity-resolution and graph neural-network learning across the 11 joinable tables via well_id + production_date.

---

Loading

from datasets import load_dataset
ds = load_dataset("xpertsystems/oil013-sample", data_files="daily_production.csv")
print(ds["train"][0])

Or with pandas:

import pandas as pd
wells   = pd.read_csv("hf://datasets/xpertsystems/oil013-sample/wells_master.csv")
daily   = pd.read_csv("hf://datasets/xpertsystems/oil013-sample/daily_production.csv")
lift    = pd.read_csv("hf://datasets/xpertsystems/oil013-sample/artificial_lift.csv")
labels  = pd.read_csv("hf://datasets/xpertsystems/oil013-sample/production_labels.csv")

# Join daily production to wells master for asset-type / completion-type features
joined = daily.merge(wells, on="well_id")

# Join labels to daily production (monthly labels propagated to all days in month)
labels["label_date"] = pd.to_datetime(labels["label_date"])
daily["production_date"] = pd.to_datetime(daily["production_date"])

---

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 production-engineering and decline-curve ML research, not for live production-allocation decisions. A few notes:

1. Initial rates run higher than declared base parameter. The generator's --mean-initial-oil-rate-bopd parameter is 950 BOPD, but the actual observed mean is ~1100 BOPD because two stacked lognormal multipliers (basin profile qi_mult × per-well lognormal(0, 0.25)) compound to a multiplier mean > 1. Same pattern for GOR (declared 1450, observed ~1800). This is realistic (real production distributions have positive skew), but if you need a pure declared-target match, scale --mean-initial-oil-rate- bopd down by ~13% to compensate for the lognormal-compound bias.

2. Anomaly and downtime injection rates are very low. The generator divides anomaly_injection_rate / 365.25 and downtime_event_rate / 365.25 to convert per-year rates to per-day probabilities. At sample defaults (3% per year, 2.2% per year), this gives 0.0001 daily probability — essentially zero anomalies in the daily timeseries (0.01% rate observed). The downtime_events.csv table is separately generated via a Poisson model and is populated (~2 events/well), so downtime ML training uses that table, not the per-day anomaly flag.

3. Forecast class distribution is heavily skewed toward "stable" (~97%) because the simulation runs only 365 days — Arps hyperbolic decline with mean b=1.0 and di=0.23 produces only ~20-27% rate decline in year 1, which keeps most wells in the "stable" class (oil_rate ≥ 0.60 × qi). For long-horizon forecast ML, use the full product with --simulation-days 1800+ to see meaningful class diversity (steep_decline, lift_limited, water_breakthrough all populate substantially over 3-5 years).

4. Stimulation events are extremely sparse (~1 event in 250 wells at sample scale) because the generator uses a one-time Bernoulli draw per well with combined probability ~2.4%. Full product (120K wells × 3650 days) gives ~3000 stimulation events with full event-type diversity. For workover ML at sample scale, use the workover_candidate_flag in production_labels.csv (synthesized from production patterns) rather than the literal stimulation_events table.

5. Mass balance is exact (>99.99%) because the generator's cumulative_oil_bbl column is computed as a running sum of oil_rate_bopd. This is a property of the simulation, not a physics test — but it does confirm proper integration. Use it as sanity check, not as evidence of advanced reservoir physics.

6. Pressure decline is linear, not exponential. The generator uses p(d) = p0 × (1 - 0.22 × d/sim_days), which is a simple linear depletion model. Real reservoirs follow material-balance- driven decline with B-factor and aquifer support — for reservoir-engineering-grade decline modeling, use SPE-PEH-Vol-V compliant tools rather than the OIL-013 pressure column.

7. Allocation methods are uniformly weighted, not conditioned on well type or facility. Real production allocations heavily favor test-separator for low-rate wells, model-based for commingled pads, and production-logging for problem wells. Future generator v1.1 will introduce conditioning.

---

Full product

The full OIL-013 dataset ships at 120,000 wells × 3,650 days (prod mode) producing several hundred million daily production rows with substantial populated stimulation/workover events, full multi-year decline curves enabling meaningful forecast-class diversity, and basin-conditioned operator behavior priors — licensed commercially. Contact XpertSystems.ai for licensing terms.

📧 pradeep@xpertsystems.ai 🌐 https://xpertsystems.ai

---

Citation

@dataset{xpertsystems_oil013_sample_2026,
  title  = {OIL-013: Synthetic Production Time-Series Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/oil013-sample}
}

Generation details

• Sample version : 1.0.0
• Random seed : 42
• Generated : 2026-05-22 12:39:33 UTC
• Wells : 250
• Simulation days : 365
• Basins : 10 (Permian, Eagle Ford, Bakken, Marcellus, Haynesville, GoM, North Sea, Middle East, W Canada, Brazil Pre-Salt)
• Asset types : 8 (unconventional shale oil, tight oil, shale gas, deepwater, offshore sandstone, carbonate, heavy oil, deepwater carbonate)
• Completion types : 6 (horizontal multistage frac, vertical, deviated, multilateral, open hole, cased hole)
• Lift types : 6 (natural flow, ESP, rod pump, gas lift, PCP, plunger lift)
• Downtime types : 8 (planned, unplanned, facility, weather, flow assurance, lift, integrity, power)
• Flow assurance : 6 (scale, paraffin, hydrate, sand, emulsion, corrosion)
• Forecast classes : 6 (stable, moderate decline, steep decline, water breakthrough, lift limited, workover candidate)
• Calibration basis : Arps (1945), SPE PEH Vol V, SPE 152596, SPE 167242, SPE 174021, API RP-11ER, EIA AEO, Rystad ShaleWellCube, IHS Markit, NACE, IOGP allocation
• Overall validation: 100.0/100 — Grade A+