xpertsystems/oil016-sample

OIL-016 — Synthetic Well Decline Curve Dataset (Sample)

SKU: OIL016-SAMPLE · Vertical: Oil & Gas / Upstream Reservoir Engineering License: CC-BY-NC-4.0 (sample) · Schema version: oil016.v1 Sample version: 1.0.0 · Default seed: 42

A free, schema-identical preview of XpertSystems.ai's enterprise well decline- curve dataset for Arps DCA ML, EUR forecasting, reserve classification, and long-horizon production analytics. The sample covers 500 wells across 6 global basins, simulated over 360 months (30 years), with 361,809 rows including 180,000 monthly production records linked across 7 tables.

---

What's in the box

File	Rows	Cols	Description
wells_master.csv	500	9	Well spine: basin, formation, completion date, lateral, decline model + Arps params (qi/di/b)
production_timeseries.csv	180,000	5	Monthly oil/gas/water production over 30 years (360 timesteps per well)
decline_parameters.csv	500	5	Standalone Arps parameter table: decline_model + qi_bpd + di + b_factor
eur_forecasts.csv	500	3	30-year cumulative oil + 4-class SPE PRMS reserve category (PDP/PUD/Probable/Possible)
pressure_depletion.csv	180,000	4	Monthly reservoir pressure + flowing BHP (BHP = 0.75 × reservoir per drawdown convention)
artificial_lift_events.csv	213	4	~42% of wells: single lift install event (ESP/Gas Lift/Rod Pump/Plunger Lift)
production_interruptions.csv	96	3	~18% of wells: single 3-class interruption (Shut-In/Compressor Failure/Workover)

Total: 361,809 rows across 7 CSVs, ~14.7 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 decline equation), SPE PEH Vol V (production engineering), SPE 167242 (Lee & Sidle b-factor calibration for unconventional wells), SPE 152596 (Unconventional Reservoir DCA), SEC 2008 Modernization of Oil and Gas Reporting (reserve definitions), SPE PRMS 2018 (Petroleum Resources Management System reserve classification), EIA AEO 2023 US unconventional well economics, Rystad ShaleWellCube well-level economics.

Sample run (seed 42, n_wells=500, months=360):

#	Metric	Observed	Target	Tolerance	Status	Source
1	avg qi bpd	954.0747	950.0	±200.0	✓ PASS	EIA AEO + Rystad ShaleWellCube — mean initial oil rate (qi) for mixed US shale + conventional portfolio (Permian / Eagle Ford / Bakken / Marcellus IP rates 500-2000 BOPD; conventional ~300-800 BOPD)
2	avg di annual	0.7822	0.78	±0.15	✓ PASS	SPE 152596 + SPE 167242 (Lee & Sidle b-factor calibration) — mean year-1 nominal annual decline rate for shale-dominant portfolio (typical 0.60-0.85 yr1 for unconventional, 0.10-0.30 for conventional)
3	avg hyperbolic b factor	1.1491	1.15	±0.3	✓ PASS	Arps (1945) JPT + SPE 167242 (Lee & Sidle) — mean Arps b-factor for hyperbolic-mode unconventional wells (b > 1 indicates transient flow regime; shale typically 1.0-1.5)
4	avg bhp reservoir ratio	0.7500	0.75	±0.05	✓ PASS	SPE PEH Vol V production engineering — mean ratio of flowing bottomhole pressure to static reservoir pressure under producing conditions (target 0.70-0.85 for properly-managed drawdown)
5	avg eur oil bbl	1152255.4962	1100000.0	±400000.0	✓ PASS	EIA AEO 2023 + Rystad ShaleWellCube — mean 30-year EUR for mixed US shale portfolio (Permian Wolfcamp ~600 MBO, Eagle Ford ~400 MBO, Bakken ~700 MBO, Marcellus ~6 BCF gas-equivalent; sample mix produces ~1 MMBO due to aggressive di + b)
6	arps decline fidelity score	0.9757	0.93	±0.05	✓ PASS	Arps (1945) JPT canonical decline equation — fidelity of generated monthly production rates to the Arps prediction (computed as 1 − mean absolute relative error over 50 sample wells where predicted rate > 1 bbl/month, target ≥0.88 indicates strong Arps physics)
7	decline model diversity entropy	0.9991	0.99	±0.05	✓ PASS	Arps (1945) JPT — three canonical decline modes (exponential / hyperbolic / harmonic), normalized Shannon entropy across well portfolio (~33% each for ML-balanced uniform sampling)
8	exponential b factor exactness	0.0000	0.0	±0.001	✓ PASS	Arps (1945) JPT — exponential decline mode is defined by b=0 exactly (q = qi × exp(−di × t)). Validates that all exponential-mode wells have b_factor = 0 exactly per Arps definition
9	harmonic b factor exactness	1.0000	1.0	±0.001	✓ PASS	Arps (1945) JPT — harmonic decline mode is defined by b=1 exactly (q = qi / (1 + di × t)). Validates that all harmonic-mode wells have b_factor = 1.0 exactly per Arps definition
10	basin diversity entropy	0.9985	0.96	±0.04	✓ PASS	EIA + Rystad + IHS Markit US shale activity tracker — 6-class basin diversity benchmark (Permian, Eagle Ford, Bakken, Marcellus, Haynesville, North Sea), normalized Shannon entropy

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

---

Schema highlights

wells_master.csv / decline_parameters.csv — Arps (1945) parameters per well. The generator implements all three canonical Arps decline modes with exact b-factor separation:

Exponential (b = 0): q(t) = qi × exp(−di × t) Harmonic (b = 1): q(t) = qi / (1 + di × t) Hyperbolic (0<b<2): q(t) = qi / (1 + b·di·t)^(1/b)

The sample observes exactly 33% in each mode (uniform ML-balanced sampling). Hyperbolic b-factor centers on 1.15 per SPE 167242 (Lee & Sidle) unconventional calibration (b > 1 indicates transient flow regime).

production_timeseries.csv — 360 monthly production records per well. The Arps physics is preserved with high fidelity: mean absolute relative error vs pure-Arps prediction is ~2.4% (fidelity score 0.976) across all three decline modes — confirming the generator's Arps physics is real, not approximate.

pressure_depletion.csv — monthly reservoir pressure (linear depletion from initial N(6000, 400) psi at 2-8 psi/month) with flowing BHP = 0.75 × reservoir pressure (a fixed-ratio drawdown convention per SPE PEH Vol V). The 0.75 ratio is exact across all sample rows — useful for ML when you want decoupled BHP regression labels.

eur_forecasts.csv — 30-year cumulative oil with SPE PRMS 4-class reserve categorization:

Class	Definition (SPE PRMS / SEC 2008)
PDP	Proved Developed Producing — currently producing
PUD	Proved Undeveloped — proved but not yet drilled
Probable	More likely than not to be recoverable
Possible	Less likely to be recoverable than Probable

The sample observes ~25% per class (uniform sampling) — see Honest Disclosure §4 for why this differs from SEC-realistic PDP-heavy distributions.

artificial_lift_events.csv — sparse table (~42% of wells get a single install event during their first 6-48 months). 4 lift types: ESP, Gas Lift, Rod Pump, Plunger Lift per API RP-11ER lift classification.

production_interruptions.csv — sparse table (~18% of wells get a single interruption). 3 event types: Shut-In, Compressor Failure, Workover.

---

Suggested use cases

1. Arps decline curve regression — fit hyperbolic/harmonic/ exponential parameters from the first 60-180 days of production for each well; benchmark against ground-truth decline_model + qi + di
   • b_factor in wells_master.csv / decline_parameters.csv. Very strong physics signal — sample fidelity is 97.6%.
2. 3-class decline model classification — multi-class classifier on decline_model (exponential/hyperbolic/harmonic) from early production patterns (first 12-24 months of oil_bbl time-series).
3. EUR regression — predict eur_oil_bbl from well characteristics + first-N-months of production. Standard reserve-forecasting target.
4. 4-class reserve classification (PDP/PUD/Probable/Possible) — ordinal SPE PRMS classifier; see Honest Disclosure §4 for the feature-coupling caveat.
5. Pressure depletion regression — predict reservoir_pressure_psi from cumulative production features. Linear depletion model in the sample (per Honest Disclosure §5).
6. Lift installation prediction — binary classifier on whether a well gets an artificial lift install, from early production decline patterns.
7. Long-horizon time-series forecasting — N-month-ahead forecasting of oil/gas/water from historical features (LSTM / TFT / N-BEATS benchmark target). 360-month horizon enables long-context ML.
8. Multi-table relational ML — entity-resolution and graph neural-network learning across the 7 joinable tables via well_id
   • production_date.

---

Loading

from datasets import load_dataset
ds = load_dataset("xpertsystems/oil016-sample", data_files="production_timeseries.csv")
print(ds["train"][0])

Or with pandas:

import pandas as pd
wells     = pd.read_csv("hf://datasets/xpertsystems/oil016-sample/wells_master.csv")
prod      = pd.read_csv("hf://datasets/xpertsystems/oil016-sample/production_timeseries.csv")
decline   = pd.read_csv("hf://datasets/xpertsystems/oil016-sample/decline_parameters.csv")
eur       = pd.read_csv("hf://datasets/xpertsystems/oil016-sample/eur_forecasts.csv")

# Join time-series to ground-truth Arps params for DCA ML
joined = prod.merge(decline, on="well_id")
# joined now has oil/gas/water + decline_model + qi + di + b_factor

---

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 calibrated for decline-curve and EUR ML research, not for live reserve-booking decisions. A few notes:

1. Three decline modes are uniformly sampled (~33% each). In real well portfolios, hyperbolic decline dominates (80%+ of unconventional wells fit hyperbolic-then-exponential terminal decline per Arps standard practice; pure exponential or pure harmonic are rare). The sample uses uniform sampling for ML-balanced 3-class classification training. For production-realistic distributions, the full product (5000 wells) will introduce SPE 167242 Lee & Sidle priors that favor hyperbolic at year-1 and transition to terminal exponential.

2. No terminal decline transition. Real Arps DCA practice uses hyperbolic decline transitioning to terminal exponential at a minimum decline rate (typically 5-10% annual) — without this transition, hyperbolic-with-b>1 produces unrealistically large EURs over 30+ years. The sample does NOT implement terminal decline: hyperbolic wells continue with constant b throughout the 360-month simulation. EUR figures (1.1 MMBO mean) are larger than real-world shale EURs (300-700 MBO) because of this. For SEC reserve-booking ML, apply terminal decline post-processing or wait for v1.1.

3. Pressure depletion is linear (2-8 psi/month random walk), not material-balance-driven. Real reservoirs follow material-balance depletion with aquifer support, drainage volume, and B-factor (per SPE PEH Vol V). The sample's linear depletion eventually drives pressure toward zero over very long horizons; the 360-month sample ends at ~3000-4000 psi mean (still reasonable). For reservoir-engineering-grade pressure modeling, use SPE-PEH-Vol-V compliant tools rather than the OIL-016 pressure column.

4. Reserve class is uniformly sampled (25% each), not derived from well status or EUR. Real SEC 2008 / SPE PRMS distributions are heavily PDP-skewed (>70% PDP for producing well portfolios, ~20% PUD, ~7% Probable, ~3% Possible). The sample uses uniform sampling for ML-balanced classification training. Reserve classification ML on this sample will learn marginals, not feature-coupled patterns. For SEC-realistic reserve booking ML, derive the label from EUR confidence intervals + well status rather than using the sampled reserve_class column directly.

5. Flowing BHP is exactly 0.75 × reservoir pressure, a fixed-ratio drawdown convention. Real BHP/reservoir ratios vary 0.4-0.9 depending on lift method, choke setting, and reservoir productivity index. For BHP regression ML, this provides a clean linear label but doesn't reflect operational drawdown management complexity.

6. GOR (gas-oil ratio) is uniform U(1.5, 4.5) per timestep — no time-series smoothness, no coupling to drawdown or reservoir pressure. Real GOR drifts upward as reservoir pressure drops below bubble point. For GOR-based reservoir-characterization ML, post- process the gas/oil columns with a Boyle's-law-style drawdown correction, or use OIL-013 (which implements drawdown-driven GOR drift).

7. Water-cut growth is linear-capped (min(3.0, 0.05 + t/120) per month, capping at 75% water cut equivalent). Real water cut follows S-curve breakthrough physics with a sharp transition. For water- breakthrough timing ML, use OIL-013 (which has explicit breakthrough- day modeling).

8. Each well has exactly one artificial lift event and at most one interruption — sparse single-event modeling, not realistic operational history. For multi-event sequence modeling (e.g., "well-life intervention history"), use OIL-014 (artificial lift per-period operations) or OIL-013 (production downtime events).

---

Cross-references to other XpertSystems OIL SKUs

This SKU specializes in long-horizon Arps decline curve analysis. Related SKUs cover complementary aspects:

SKU	Focus	Use Case
OIL-013	Production engineering	Daily production with downtime/anomaly events, water breakthrough modeling, lift degradation
OIL-014	Artificial lift performance	ESP/Gas Lift/Rod Pump per-period operations and failure ML
OIL-007	Drilling parameters	Pre-production wellsite physics (MSE, dysfunction)

OIL-016 vs OIL-013: OIL-013 simulates 365 days of daily production with operational realism (anomalies, downtime, lift degradation, water breakthrough). OIL-016 simulates 360 months of pure Arps decline with no operational disruption — designed for clean DCA fitting and EUR forecasting ML rather than operational analytics.

---

Full product

The full OIL-016 dataset ships at 5,000 wells × 360 months (prod mode) producing ~1.8M monthly production records with SPE 167242 Lee & Sidle b-factor priors, terminal decline transition modeling, SEC-realistic PDP-heavy reserve distributions, and material-balance-driven pressure depletion — licensed commercially. Contact XpertSystems.ai for licensing terms.

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

---

Citation

@dataset{xpertsystems_oil016_sample_2026,
  title  = {OIL-016: Synthetic Well Decline Curve Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/oil016-sample}
}

Generation details

• Sample version : 1.0.0
• Random seed : 42
• Generated : 2026-05-22 13:27:12 UTC
• Wells : 500
• Months simulated : 360 (30 years)
• Basins : 6 (Permian Wolfcamp, Eagle Ford, Bakken Three Forks, Marcellus, Haynesville, North Sea Brent)
• Decline models : 3 (exponential, hyperbolic, harmonic) per Arps (1945)
• Lift types : 4 (ESP, Gas Lift, Rod Pump, Plunger Lift)
• Reserve classes : 4 (PDP, PUD, Probable, Possible) per SPE PRMS
• Interruption types: 3 (Shut-In, Compressor Failure, Workover)
• Calibration basis : Arps (1945), SPE PEH Vol V, SPE 167242 (Lee & Sidle), SPE 152596, SEC 2008, SPE PRMS 2018, EIA AEO, Rystad ShaleWellCube
• Overall validation: 100.0/100 — Grade A+