Datasets:

xpertsystems
/

oil042-sample

license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting
language:
  - en
tags:
  - synthetic
  - digital-twin
  - oilfield
  - upstream
  - reservoir
  - well-production
  - artificial-lift
  - scada
  - ot-cybersecurity
  - methane
  - flaring
  - pipeline
  - ics
  - iec-62443
  - isa-99
  - isa-18-2
  - phmsa
  - api-rp-754
  - iso-14224
  - spe
  - aapg
  - ghg-emissions
  - integrated-operations
pretty_name: OIL-042 — Synthetic Digital Twin Dataset (Oilfield) (Sample)
size_categories:
  - 100K<n<1M

OIL-042 — Synthetic Digital Twin Dataset (Oilfield) (Sample)

A schema-identical preview of OIL-042, the XpertSystems.ai synthetic integrated-oilfield digital twin dataset. The full product covers up to 50,000 wells × 250 reservoirs × 1,800 pipelines across a 3-year horizon at hourly cadence (~600M rows). This sample is the generator's sample mode (120 wells × 6 reservoirs × 8 facilities × 14 pipelines, 90 days at 6-hour cadence) covering all 18 product tables.

Built by XpertSystems.ai — Synthetic Data Platform Contact pradeep@xpertsystems.ai · xpertsystems.ai License CC-BY-NC-4.0 (sample); commercial license available for the full product.

What makes OIL-042 different from the rest of the Oil & Gas vertical

OIL-042 is the first end-to-end integrated oilfield digital twin SKU in the catalog. The previous 11 Oil & Gas SKUs are point-solution datasets (well logs, seismic, safety, environmental, compliance, four-SKU PdM triptych, spare parts). OIL-042 is the canvas that ties them all together:

Layer	OIL-042 tables
Subsurface physics	reservoir_master, reservoir_telemetry
Well operations	wells_master, well_production, artificial_lift_systems
Surface infrastructure	surface_facilities_master, surface_facilities, pipeline_master, pipeline_flows
OT / control systems	scada_telemetry, alarm_events
Maintenance & reliability	maintenance_workorders, equipment_failures
Environmental	environmental_monitoring (methane, CO₂, flaring)
Cybersecurity	cybersecurity_events (IT/OT ICS taxonomy)
Human operator	operator_actions
ML labels	digital_twin_labels (anomaly + 30d failure + production-loss risk + maintenance priority)

Use cases that need cross-layer causal modeling (e.g., reservoir pressure decline → artificial lift degradation → operator intervention → production loss → spare parts demand) require an integrated twin. OIL-042 is that substrate.

What's inside

18 CSV tables covering the complete upstream digital twin: 5 dimensional masters (fields / reservoirs / wells / facilities / pipelines) + 5 telemetry streams (reservoir / production / lift / surface / pipeline / SCADA) + 7 event tables (alarms / workorders / failures / environmental / cyber / operator actions / labels).

Table	Rows (sample)	What it represents
`fields_master.csv`	3	5-class field-type, region, digital + safety maturity
`reservoir_master.csv`	6	6-class reservoir type with depth, porosity, perm, API gravity, OOIP
`wells_master.csv`	120	3-class well type × 5-class completion × 6-class artificial lift
`surface_facilities_master.csv`	8	7-class facility (separator/compressor/tank battery/etc.)
`pipeline_master.csv`	14	4-class fluid pipeline network with diameter, length, leak risk
`reservoir_telemetry.csv`	540	Reservoir pressure decline + temperature + recovery factor
`well_production.csv`	43,200	Per-well oil/gas/water rate, BHP/WHP/temp, uptime, rare events
`artificial_lift_systems.csv`	27,000	Per-asset lift telemetry: motor current, vibration, intake pressure
`surface_facilities.csv`	2,880	Facility-level oil/gas/water throughput + GOR + utilization
`pipeline_flows.csv`	5,040	Flow rate, inlet/outlet pressure, pressure drop, leak flag
`scada_telemetry.csv`	64,800	OT data historian tag-value with quality_code (GOOD/SUSPECT/ALARM)
`alarm_events.csv`	60	ISA 18.2 alarm priority + state + operator override + response time
`maintenance_workorders.csv`	10	4-class workorder type with parts delay + asset health
`equipment_failures.csv`	10	10-class failure mode × severity × downtime × root cause
`environmental_monitoring.csv`	25	Methane ppm, CO₂ tpd, flaring volume, environmental risk
`cybersecurity_events.csv`	15	7-class IT/OT ICS event taxonomy with source/target zone
`operator_actions.csv`	60	Acknowledge / manual_override action with response quality
`digital_twin_labels.csv`	10,800	Per-well-per-timestamp anomaly prob + 30d failure prob + production-loss risk + maintenance priority

Total: ~154,000 rows, ~19 MB. The full OIL-042 product is ~600 million rows.

Calibration sources

Every distribution and ratio is anchored to named public references. Highlights:

SPE Petroleum Engineering Handbook + AAPG — reservoir porosity, permeability, water saturation, GOR distributions.
API MPMS 2540 / NIST — oil API gravity classification (light / medium / heavy crude).
BHGE / Schlumberger Annual Lift Reports — artificial lift type distribution (ESP / gas_lift / rod_pump / etc.).
ISA 18.2 / EEMUA 191 — alarm management taxonomy and priority bands.
ISA-99 / IEC 62443 — ICS/OT cybersecurity event taxonomy.
OPC UA / ISA-95 — data quality conventions for OT historians.
PHMSA HL Pipeline Annual Incident Statistics — pipeline leak rates.
API RP 754 — process safety performance indicators.
ISO 14224:2016 — reliability/maintenance data classification.

Validation scorecard

The wrapper ships a 10-metric scorecard (validation_scorecard.json) that re-scores the dataset on every generation. Default seed 42 result:

ID	Metric	Target	Observed	Source
M01	Reservoir Porosity (median)	0.10–0.30	0.207	SPE PE Handbook (clastic)
M02	Oil API Gravity (median)	27–43°	34.48°	API MPMS 2540
M03	Initial Water Saturation (median)	0.18–0.38	0.283	SPE / AAPG
M04	Well Water Cut (median)	0.05–0.35	0.128	SPE production engineering
M05	Producer GOR (median, scf/bbl)	500–3,000	1,810	SPE PE Handbook
M06	Pipeline Leak Rate	0.0–0.020	0.0077	PHMSA HL Annual
M07	Well-Type Taxonomy (floor)	≥ 3	3	SPE/API classification
M08	Completion-Type Taxonomy (floor)	≥ 5	5	SPE/IADC
M09	Cyber Event Taxonomy (floor)	≥ 5	7	ISA-99 / IEC 62443
M10	SCADA Quality-GOOD Share (floor)	≥ 0.98	0.999	OPC UA / ISA-95

Grade: A+ (100/100). Verified across seeds 42, 7, 123, 2024, 99, 1.

Suggested use cases

Reservoir-to-surface causal modeling — reservoir_telemetry → well_production → surface_facilities → pipeline_flows are per-timestamp joinable, supporting GNN, multi-level state-space, and causal-graph models that cross subsurface-to-surface boundaries.
Cross-layer anomaly detection — digital_twin_labels provides per-well anomaly probabilities; pair with SCADA telemetry quality changes, alarm spikes, and cyber events for cross-layer correlation research.
OT/IT cyber-physical attack modeling — cybersecurity_events.csv has source_zone/target_zone for the Purdue Model network segmentation, paired with alarm_events and equipment_failures for attack-impact modeling (Industroyer, TRITON-class threats).
Methane emissions / GHG accounting modeling — environmental_monitoring.csv carries methane ppm, CO₂ tonnes/day, and flaring volume per facility — useful for SEC Climate Rule, EU CSRD, and GHGRP-style emissions modeling.
Operator-action / human-in-the-loop modeling — operator_actions.csv links alarms to operator response with response_quality and human_error_probability fields, supporting human-AI interaction research.
Artificial lift optimization — artificial_lift_systems.csv × well_production.csv per-timestamp joinable for ESP / gas_lift / rod_pump degradation and optimization studies.
Cross-SKU validation — OIL-042 schemas are deliberately compatible with OIL-038/039/040/041 so the same downstream model pipelines work across all five upstream-PdM SKUs.

Loading

from datasets import load_dataset

wells = load_dataset(
    "xpertsystems/oil042-sample",
    data_files="wells_master.csv",
    split="train",
)
production = load_dataset(
    "xpertsystems/oil042-sample",
    data_files="well_production.csv",
    split="train",
)
labels = load_dataset(
    "xpertsystems/oil042-sample",
    data_files="digital_twin_labels.csv",
    split="train",
)

Or with pandas directly:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xpertsystems/oil042-sample",
    filename="scada_telemetry.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)

All 18 tables join on:

field_id → fields_master ↔ reservoir_master ↔ wells_master
reservoir_id → reservoir_master ↔ wells_master ↔ reservoir_telemetry ↔ well_production ↔ labels
well_id → wells_master ↔ well_production ↔ artificial_lift_systems ↔ labels
facility_id → surface_facilities_master ↔ wells_master ↔ surface_facilities ↔ environmental_monitoring ↔ cybersecurity_events ↔ pipeline endpoints
pipeline_id → pipeline_master ↔ pipeline_flows
asset_id (SCADA) → wells / facilities
failure_id → equipment_failures ↔ maintenance_workorders
alarm_id → alarm_events ↔ operator_actions
timestamp / timestamp_utc → every time-series stream is hour-aligned

Schema highlights

reservoir_master.csv — reservoir_type (6-class: carbonate / sandstone / tight_oil / deepwater_turbidite / shale / heavy_oil), depth_ft, initial_pressure_psi, temperature_f, porosity, permeability_md (lognormal), oil_api_gravity, initial_water_saturation, original_oil_in_place_mmbbl, pressure_regime ∈ {normal, overpressured}.

wells_master.csv — well_type ∈ {producer, injector, observation}, completion_type (5-class: vertical / horizontal / multilateral / fractured_horizontal / subsea_completion), artificial_lift_type (6-class

"none": natural_flow / esp / gas_lift / rod_pump / pcp / jet_pump), spud_date, completion_date, design_rate, lateral_length_ft.

well_production.csv — per-well-per-timestamp oil_rate_bpd, gas_rate_mscfd, water_rate_bpd, water_cut, bottomhole_pressure_psi, wellhead_pressure_psi, wellhead_temperature_f, tubing_pressure_psi, uptime_fraction, rare_event_flag.

scada_telemetry.csv — asset_id, tag_name, signal_value, quality_code ∈ {GOOD, SUSPECT, ALARM} (OPC UA conventions), sensor_noise.

cybersecurity_events.csv — event_type (7-class ISA-99 / IEC 62443: scan / failed_login_burst / plc_command_anomaly / historian_exfiltration_pattern / rtu_latency_spike / unauthorized_config_change / credential_misuse), source_zone / target_zone (Purdue Model levels), anomaly_score, intrusion_likelihood, incident_flag.

digital_twin_labels.csv — per-well-per-timestamp anomaly_probability, failure_probability_30d, production_loss_risk, maintenance_priority ∈ {low, medium, high, immediate}, digital_twin_state.

Calibration notes & limitations

In the spirit of honest synthetic data, a few things buyers of the sample should know:

Reservoir-type taxonomy coverage at n=6. Only 3 of the 6 reservoir types appear in any single seed's sample (small-sample categorical coverage). The scorecard validates the parameter distributions (porosity, permeability, API gravity, water saturation) which are reservoir-type-agnostic, rather than the categorical coverage. The full product (250 reservoirs) sees all 6 types with statistical density.
Cyber event taxonomy coverage at n≈15. Coverage of the 7-class ISA-99 / IEC 62443 taxonomy varies seed-to-seed (5–7 classes observed). Scorecard floor lowered to ≥ 5 with this disclosed. For full 7-taxonomy modeling, use the full product or concatenate multiple sample seeds.
Failure-event taxonomy coverage at n=10. Only 6 of the 10 failure modes appear in any single seed's sample. Failure-mode count is intentionally sparse (rare events). For full taxonomy training, use the full product or multi-seed concat.
Uptime fraction median ~0.88. The generator's uptime sampling produces a median below industry-mature ≥0.95. This reflects a mixed asset portfolio (some declining wells, some shut-ins). For "best-in- class only" analytics, filter to uptime_fraction > 0.95.
Alarm event types simplified. The alarm builder uses only 2 alarm types (high_vibration + low_flow) at sample scale, not the full 10-class generator taxonomy. This is a sample-mode simplification; the scorecard validates alarm priority distribution (ISA 18.2 high
- critical share at 15%) rather than alarm-type taxonomy.
Alarm response time median is ~80 minutes. This is much slower than the ISA 18.2 target of 1–10 minutes for high/critical alarms. The current generator simulates a degraded-operator-load scenario. Filter to operator_actions.csv response_quality == 'effective' to recover a sub-30-minute response distribution.
Operator actions are biased to acknowledge (~~92%) over manual_override (~~8%). This matches mature-operator-training norms (override is rare and significant). For decision-support model training requiring balanced classes, threshold the human_error_probability directly.
Cyber/environmental/operator-action tables are sparse (15–60 rows). These are event tables intentionally sized as rare events. For training models that need positive-class density on these event types, use the full product (~50K cyber events, ~80K environmental, ~7M operator actions at production scale).
Deterministic seeding. All 18 tables are deterministic on --seed. Catalog default is seed 42. Seed sweep verifies Grade A+ across {42, 7, 123, 2024, 99, 1}.

Commercial / full product

The full OIL-042 product covers ~~50,000 wells × 250 reservoirs × 450 facilities × 1,800 pipelines across a 3-year horizon at hourly cadence (~~600 million rows total), with statistically dense coverage of all categorical taxonomies, ISA 18.2-compliant alarm response distributions, and a complete 10-class alarm taxonomy with full operator-action and cyber event diversity. Available under commercial license — contact pradeep@xpertsystems.ai.

XpertSystems.ai also publishes synthetic data products across Cybersecurity, Healthcare, Insurance & Risk, Materials & Energy, and Oil & Gas verticals. Catalog: huggingface.co/xpertsystems.