xpertsystems/oil017-sample

OIL-017 — Synthetic Water Injection Dataset (Sample)

SKU: OIL017-SAMPLE · Vertical: Oil & Gas / Upstream Enhanced Oil Recovery License: CC-BY-NC-4.0 (sample) · Schema version: oil017.v1 Sample version: 1.0.0 · Default seed: 42

A free, schema-identical preview of XpertSystems.ai's enterprise water injection / waterflood dataset for sweep efficiency ML, breakthrough prediction, conformance optimization, and Voidage Replacement Ratio (VRR) analytics. The sample covers 80 injectors with 359 injector-producer connectivity links across 8 global basins and 5 flood pattern types, simulated over 3,650 days (10 years), with 136,579 rows linked across 11 tables.

---

What's in the box

File	Rows	Cols	Description
injection_wells.csv	80	20	Injector spine: basin, pattern, perm/poro/thickness, mobility ratio, injection rate, water source
connectivity_matrix.csv	359	13	Injector-producer pairs: distance, connectivity score, communication delay, fracture-assisted flag
producer_response.csv	43,798	12	Per-link-per-timestep oil/water rate + water cut + logistic post-breakthrough water rise
reservoir_pressure.csv	43,798	9	Per-link pressure timeseries + VRR + pressure support efficiency
sweep_efficiency.csv	43,798	10	Areal + vertical + displacement sweep over time, mobility-ratio-modulated per Buckley-Leverett
breakthrough_events.csv	359	9	One event per link: severity + post-breakthrough water cut + channeling-suspected flag
reservoir_labels.csv	359	9	ML labels: 4-class sweep quality grade (A/B/C/D) + 3-class flood efficiency + 3-class breakthrough risk + recovery factor
injection_profiles.csv	376	8	Multi-layer Dirichlet allocation per injector (sums to 100%) + per-layer injectivity index
water_quality.csv	3,280	10	Quarterly water samples: salinity / sulfate / hardness / bacteria / scaling risk per NACE SP0192
conformance_events.csv	13	9	Sparse channeling events: 6-class treatment (gel polymer / profile mod / selective shutoff / polymer flood / low salinity) + effectiveness
production_forecasts.csv	359	9	Per-link 5-year EUR + recovery factor + terminal water cut + economic limit flag

Total: 136,579 rows across 11 CSVs, ~14.1 MB on disk.

---

Calibration: industry-anchored, honestly reported

Validation uses a 10-metric scorecard with targets sourced exclusively to named industry standards: Buckley-Leverett (1942) fractional flow theory, Welge (1952) displacement efficiency, Dykstra-Parsons (1950) heterogeneity index, Craig (1971) SPE Monograph 3 "Reservoir Engineering Aspects of Waterflooding" (the canonical waterflood reference), SPE PEH Vol V (production engineering), SPE 14529 (Voidage Replacement Ratio guidance), Willhite (1986) "Waterflooding" SPE textbook, SPE 13855 (sweep efficiency models), Caudle (1968) flood pattern geometries, NACE SP0192 (oilfield water quality for injection), Rystad Energy + IHS Markit global waterflood operations tracker.

Sample run (seed 42, n_injectors=80, simulation_days=3,650):

#	Metric	Observed	Target	Tolerance	Status	Source
1	avg voidage replacement ratio	0.9799	0.98	±0.1	✓ PASS	SPE 14529 (Voidage Replacement Ratio guidance) + SPE PEH Vol V — mean VRR for mature waterflood portfolio (target 0.95-1.05 for pressure maintenance; <0.90 indicates under-injection, >1.10 over-injection)
2	avg mobility ratio	1.1499	1.15	±0.4	✓ PASS	Buckley-Leverett (1942) fractional flow theory + Craig (1971) SPE Monograph 3 — mean end-point mobility ratio for mixed sandstone/carbonate/heavy-oil portfolio (M < 1 favorable, M > 1 unfavorable; typical 0.5-2.5)
3	avg reservoir pressure psi	3779.8925	3850.0	±600.0	✓ PASS	SPE PEH Vol V + IHS Markit global waterflood tracker — mean reservoir pressure for mature waterflood portfolio (typical 2500-5500 psi at injection-supported equilibrium)
4	avg injection rate bwpd	14235.8650	13000.0	±4000.0	✓ PASS	Rystad Energy + IHS Markit + SPE PEH Vol V — mean water injection rate for mixed global waterflood operations (typical 5,000-25,000 BWPD per injector for moderately-thick reservoirs)
5	avg water salinity ppm	84378.8965	85000.0	±25000.0	✓ PASS	NACE SP0192 (oilfield water quality for injection) + SPE 14529 — mean water salinity for mixed produced-water + seawater portfolio (formation water 50,000-200,000 ppm TDS; seawater ~35,000 ppm)
6	avg composite sweep efficiency	0.8199	0.75	±0.15	✓ PASS	Craig (1971) SPE Monograph 3 + Welge (1952) — mean composite sweep efficiency (geometric mean of areal × vertical × displacement) for mature waterflood portfolio (0.45-0.75 typical at flood maturity)
7	connectivity delay pearson correlation	-0.8096	-0.7	±0.2	✓ PASS	Craig (1971) + SPE 13855 sweep efficiency models — expected strong inverse correlation between injector-producer connectivity score and communication delay (physics: high connectivity = short delay = fast breakthrough propagation)
8	mobility displacement pearson correlation	-0.9530	-0.85	±0.15	✓ PASS	Buckley-Leverett (1942) + Craig (1971) — expected strong inverse correlation between mobility ratio (M) and displacement efficiency (E_d ∝ 1/M^0.35 per fractional flow theory; M > 1 = unfavorable mobility = reduced displacement)
9	injection profile completeness	1.0000	1.0	±0.02	✓ PASS	SPE production allocation guidelines + IOGP injection profile standards — per-injector multi-layer injection allocations must sum to 100% (validates Dirichlet sampling produces complete profiles)
10	flood pattern diversity entropy	0.9353	0.91	±0.06	✓ PASS	Caudle (1968) flood pattern geometries + Craig (1971) SPE Monograph 3 — 5-class flood-pattern diversity benchmark (5-spot, 9-spot, line drive, peripheral, inverted 5-spot; 5-spot dominant per industry default weights [0.34, 0.18, 0.20, 0.18, 0.10]), normalized Shannon entropy

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

---

Schema highlights

injection_wells.csv — injector spine with basin-conditioned heterogeneity priors per Craig (1971):

heterogeneity_index = basin_base + N(0, 0.08) clip(0.05, 0.85) mobility_ratio = lognormal(log(1.15), 0.30) clip(0.35, 4.5) injection_pressure = reservoir_P + N(850, 270) clamped to (P_res + 80, P_frac − 120)

Injection pressure is always between reservoir pressure and fracture pressure — preventing physically-impossible "fractured injection above parting pressure" scenarios.

connectivity_matrix.csv — injector-producer pairs with Poisson- sampled producers per injector (mean 4, pattern-conditioned). Communication delay follows the canonical Craig (1971) relation:

delay = base_delay × (distance / 1800 ft)^0.25 / max(connectivity, 0.34)

The connectivity↔delay Pearson correlation is r ≈ −0.81 in the sample — strong inverse coupling confirms Craig (1971) physics.

producer_response.csv — per-link-per-timestep production with logistic post-breakthrough water rise:

Pre-breakthrough: water_cut = 18 + 20·(t/T_total) + N(0, 1.5) Post-breakthrough: water_rise = 1 / (1 + exp(−3.2·(post/720 − 0.5))) water_cut = 24 + (max_WC − 24) × water_rise

This is Buckley-Leverett-style sigmoid breakthrough behavior — slow ramp-up, steep rise around mid-breakthrough, asymptotic approach to terminal water cut. The breakthrough_flag column is the canonical binary classification target.

reservoir_pressure.csv — VRR-driven pressure depletion + injection support per SPE 14529:

depletion = 0.00012 × day × (1.05 − VRR) × P_initial support_gain = 180 × (1 − exp(−day/900)) × max(0, VRR − 0.86) pressure(t) = P_initial − depletion + support_gain + noise

VRR centered on 0.98 (target balanced injection); VRR > 0.86 produces net support gain. The sample's VRR mean is 0.980 — bullseye for SPE 14529 balanced-waterflood operations.

sweep_efficiency.csv — Buckley-Leverett / Craig (1971) sweep model with three components:

areal = 0.60 + 0.28·(1 − exp(−day/1200)) − 0.10·heterogeneity + noise vertical = 0.56 + 0.26·(1 − exp(−day/1500)) − 0.08·heterogeneity + noise displacement = 0.66 + 0.20 / mobility_ratio^0.35 − 0.04·heterogeneity

The mobility-ratio↔displacement correlation is r ≈ −0.95 — near-perfect inverse coupling per Buckley-Leverett (1942) fractional flow theory (M > 1 = unfavorable mobility = reduced displacement).

injection_profiles.csv — multi-layer Dirichlet allocation per injector with heterogeneity-driven concentration:

alpha = 1.5 + 2.0·(1 − heterogeneity) fractions ~ Dirichlet(alpha × ones(n_layers))

Per-injector fractions sum to exactly 100%. High-heterogeneity reservoirs get more spread-out (low-α) distributions; low-heterogeneity reservoirs get more even distributions.

reservoir_labels.csv — 4-class sweep grade per Welge (1952) displacement efficiency benchmarks:

Grade	Threshold (composite sweep)
A	≥ 0.72
B	0.62 ≤ sweep < 0.72
C	0.50 ≤ sweep < 0.62
D	< 0.50

---

Suggested use cases

1. Breakthrough timing regression — predict breakthrough_time_days from connectivity score + distance + heterogeneity features. Very strong physics signal: r ≈ −0.81 connectivity↔delay coupling.
2. Composite sweep efficiency regression — predict end-of-flood sweep from mobility ratio + heterogeneity features. Near-perfect physics signal: r ≈ −0.95 mobility↔displacement coupling.
3. 4-class sweep quality grade classification — ordinal classifier (A/B/C/D) on sweep grade — useful as label-only reference; see Honest Disclosure §1 for sample-scale class-imbalance caveat.
4. 3-class breakthrough risk classification — multi-class classifier (low/medium/high) from upstream operational features. Less imbalanced than sweep grade.
5. Channeling/thief zone detection — binary classifier on channeling_suspected_flag from connectivity + heterogeneity features.
6. VRR optimization — regression on voidage_replacement_ratio to identify under/over-injection scenarios per SPE 14529.
7. Scaling risk prediction — regression on scaling_risk_score from water-quality features (salinity / sulfate / hardness) per NACE SP0192.
8. Conformance treatment prediction — multi-class classifier on treatment_type from channeling/heterogeneity features (note: sparse table, see Honest Disclosure §3).
9. EUR forecasting — regression on estimated_ultimate_recovery_bbl per injector-producer link from operational history features.
10. Multi-table relational ML — entity-resolution and graph neural-network learning across the 11 joinable tables via injector_id + producer_id.

---

Loading

from datasets import load_dataset
ds = load_dataset("xpertsystems/oil017-sample", data_files="producer_response.csv")
print(ds["train"][0])

Or with pandas:

import pandas as pd
inj    = pd.read_csv("hf://datasets/xpertsystems/oil017-sample/injection_wells.csv")
conn   = pd.read_csv("hf://datasets/xpertsystems/oil017-sample/connectivity_matrix.csv")
prod   = pd.read_csv("hf://datasets/xpertsystems/oil017-sample/producer_response.csv")
sweep  = pd.read_csv("hf://datasets/xpertsystems/oil017-sample/sweep_efficiency.csv")
labels = pd.read_csv("hf://datasets/xpertsystems/oil017-sample/reservoir_labels.csv")

# Join sweep timeseries to injector mobility ratio for Buckley-Leverett ML:
sweep_full = sweep.merge(inj[["injector_id", "mobility_ratio", "heterogeneity_index"]], on="injector_id")
# Now you have displacement_efficiency ↔ mobility_ratio for sweep ML

---

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 waterflood / EOR ML research, not for live reservoir-management decisions. Several important notes:

1. Sweep quality grade is heavily skewed toward A at sample scale. The 10-year simulation horizon drives most links to mature/high sweep state (composite sweep ~0.82 mean, well above the 0.72 grade-A threshold). All ~360 sample links classify as grade A and flood_efficiency_class=high in the seed-42 run. The 4-class sweep classification task is degenerate at sample scale. Use recovery_factor_pct as a continuous regression target instead, or use the breakthrough_risk_class (which has all three classes populated: ~59% medium / ~26% high / ~14% low). The full product (45K injectors, longer / mixed simulation durations) gives proper class diversity.

2. optimization_priority is all "standard" in the sample because the "urgent" condition requires grade C/D AND severity > 0.60, and no rows trip grade C/D at this simulation maturity. Treat this column as reference-only at sample scale.

3. Conformance events are extremely sparse (~3.6% of links at sample scale, ~13 events for 359 links). Conformance treatment ML on this sample will not have enough positive examples for robust classifier training. For conformance ML at sample scale, use the channeling_suspected_flag in breakthrough_events.csv (synthesized from severity, more populated) rather than the literal conformance events table. The full product generates orders of magnitude more conformance events.

4. Breakthrough rate is ~75% by end of simulation because the 3650-day (10-year) horizon exceeds 4× the median breakthrough time (900 days). This is physically correct for mature waterfloods but means the breakthrough_flag column saturates over time. For early-breakthrough ML, filter to day < median_breakthrough_time (900 days) to study pre-breakthrough → breakthrough transitions.

5. Heterogeneity ↔ sweep coupling is weak in the sample (r ≈ −0.14 for areal sweep, r ≈ +0.10 for breakthrough time). The generator includes heterogeneity as a small additive modifier rather than a strong multiplicative driver — real Dykstra-Parsons V_DP coefficients produce much stronger heterogeneity↔sweep coupling. For heterogeneity-driven sweep ML, use the full product or post-process with V_DP recalibration.

6. All injection pressure is clamped between reservoir P + 80 and fracture P − 120 psi. This is a deliberate safety constraint (preventing physically-impossible above-parting injection) but means the injection_pressure_psi column has a tight bounded range and cannot represent over-pressure / fractured-injection scenarios. For above-parting injection ML, manually relax the clamp in the generator.

7. Mean breakthrough time is 902 days (cfg target 730) because the generator applies a (distance/1800)^0.25 scaling factor that biases delays upward for far-spaced links. The bre_delay = base_delay × (1.0 + 0.22 × heterogeneity) formula also extends delays in heterogeneous reservoirs. Both are correct physics — but means the declared mean_breakthrough_days=730 is the base parameter before distance/heterogeneity adjustment, not the actual mean.

8. Water quality scaling risk averages 0.62 — high because sulfate + hardness multipliers compound. Real scaling risk would condition on water source (seawater→high sulfate scaling, produced water→high barium/strontium). Sample uses uniform-conditioned mineralogy.

---

Full product

The full OIL-017 dataset ships at 45,000 injectors × 10-year simulation (prod mode) producing several hundred million producer- response rows with Dykstra-Parsons-calibrated heterogeneity coupling, proper 4-class sweep grade diversity (mixed simulation durations to populate D/C/B grades), realistic conformance event rates (15-25% of heterogeneous links), and water-source-conditional mineralogy — licensed commercially. Contact XpertSystems.ai for licensing terms.

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

---

Cross-references to other XpertSystems OIL SKUs

This SKU specializes in waterflood / water injection EOR analytics. Related SKUs cover complementary aspects:

SKU	Focus	Use Case
OIL-013	Production engineering	Daily production with anomaly events, water breakthrough modeling at single-well scale
OIL-014	Artificial lift performance	ESP / Gas Lift / Rod Pump operations (downstream of waterflood-aided wells)
OIL-016	Decline curve analysis	Long-horizon Arps DCA + EUR + reserve classification (without waterflood support)
OIL-017	Water injection / EOR	Injector-producer connectivity + sweep efficiency + breakthrough at field scale (this SKU)

OIL-017 vs OIL-013: OIL-013 simulates single-well daily production with operational realism. OIL-017 simulates injector-producer pair dynamics at field scale with explicit connectivity modeling, sweep efficiency physics, and VRR-driven pressure maintenance. Use OIL-013 for well-level ML, OIL-017 for field-scale waterflood optimization ML.

---

Citation

@dataset{xpertsystems_oil017_sample_2026,
  title  = {OIL-017: Synthetic Water Injection Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/oil017-sample}
}

Generation details

• Sample version : 1.0.0
• Random seed : 42
• Generated : 2026-05-22 13:38:39 UTC
• Injectors : 80
• Simulation days : 3,650 (10 years)
• Timestep : 30 days (~monthly)
• Connectivity links: 359
• Basins : 8 (Permian, North Sea, Middle East, Gulf of Mexico, Brazil Pre-Salt, Canadian Heavy Oil, ADNOC Carbonate, Kuwait Burgan)
• Lithology classes : 4 (sandstone, carbonate, deepwater sand, heavy oil sand)
• Flood patterns : 5 (5-spot, 9-spot, line drive, peripheral, inverted 5-spot per Caudle 1968)
• Water sources : 4 (produced water, seawater, aquifer, treated mixed)
• Treatment types : 6 (none, gel polymer, profile modification, selective shutoff, polymer flood, low salinity)
• Sweep grades : 4 (A, B, C, D per Welge 1952 thresholds)
• Calibration basis : Buckley-Leverett (1942), Welge (1952), Dykstra-Parsons (1950), Craig (1971) SPE Monograph 3, SPE PEH Vol V, SPE 14529, Willhite (1986), Caudle (1968), NACE SP0192, Rystad, IHS Markit
• Overall validation: 100.0/100 — Grade A+