xpertsystems/oil007-sample

--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - drilling - drilling-parameters - mwd - lwd - rop-optimization - mse - bit-wear - drilling-dysfunctions - xpertsystems pretty_name: "OIL-007 — Synthetic Drilling Parameters Dataset (Sample)" size_categories: - 100K<n<1M --- # OIL-007 — Synthetic Drilling Parameters Dataset (Sample) **SKU:** `OIL007-SAMPLE` · **Vertical:** Oil & Gas / Upstream Drilling Operations **License:** CC-BY-NC-4.0 (sample) · **Schema version:** `oil007.v1` **Generator version:** `1.0.0` · **Default seed:** `42` A free, schema-identical preview of XpertSystems.ai's enterprise drilling- parameters dataset for ROP optimization, MSE analysis, dysfunction detection, and drilling-process ML. The sample covers **100 wells** across **11 global basins** and **12 well types** with **220,871 rows** of high-cadence drilling telemetry linked across **9 tables**. --- ## What's in the box | File | Rows | Cols | Description | |---|---:|---:|---| | `wells_master.csv` | 100 | 13 | Well spine: type, basin, trajectory, rig, mud system, casing | | `drilling_timeseries.csv` | 109,870 | 10 | High-cadence per-stand ROP / WOB / torque / RPM / SPP / hook load | | `mud_properties.csv` | 21,974 | 10 | Mud weight, PV, YP, ECD, gels, mud temp | | `hydraulics_log.csv` | 21,974 | 9 | Flow, pump pressure, annular pressure, annular velocity, bit HHP | | `vibration_spectra.csv` | 54,935 | 8 | Axial / lateral / torsional g's, stick-slip index, whirl index | | `mse_log.csv` | 10,987 | 6 | Per-stand MSE (Teale formulation), bit efficiency, formation UCS | | `bit_wear_log.csv` | 344 | 9 | IADC dull grades, footage, bit size/type (PDC/RC/hybrid) | | `drilling_events.csv` | 587 | 7 | 11-class dysfunction events (stick-slip, stuck pipe, washout, mud loss, etc.) | | `drilling_labels.csv` | 100 | 6 | ML labels: optimal ROP, dysfunction risk, MSE efficiency, drilling grade | Total: **220,871 rows** across 9 CSVs, ~14.0 MB on disk. --- ## Calibration: industry-anchored, honestly reported Validation uses a **10-metric scorecard** with targets sourced exclusively to **named industry standards**: SPE 178850, SPE 96652 (Dupriest & Koederitz), Teale (1965) MSE foundational paper, API RP-7G (drill stem design), API RP-13B-1 (drilling fluids), IADC Drilling Manual, IADC dull grading, Bourgoyne et al. (1986) Applied Drilling Engineering, Rystad Energy global rig fleet, and Spears & Associates bit market reports. **Sample run** (seed `42`, n_wells=100): | # | Metric | Observed | Target | Tolerance | Status | Source | |---|---|---:|---:|---:|---|---| | 1 | avg rop ft hr | 83.8562 | 85.0 | ±25.0 | ✓ PASS | SPE 178850 + Rystad Energy — global mean ROP across mixed onshore/offshore land/platform portfolio | | 2 | avg wob klbs | 33.7322 | 32.0 | ±8.0 | ✓ PASS | API RP-7G + SPE Drilling Handbook — global mean WOB across PDC/RC mixed bit portfolio | | 3 | avg surface torque klbft | 13.7738 | 14.5 | ±4.0 | ✓ PASS | API RP-7G + SPE Drilling Engineering — global mean surface torque across mixed trajectory portfolio | | 4 | avg mud weight ppg | 11.4375 | 11.2 | ±1.5 | ✓ PASS | API RP-13B-1 + SPE drilling fluids literature — global mean mud weight across conventional/HPHT/deepwater mix | | 5 | ecd margin ppg | 0.3709 | 0.45 | ±0.15 | ✓ PASS | SPE 178850 + IADC Drilling Manual — ECD margin (ecd minus static MW) maintained during circulation | | 6 | avg mse psi | 37376.3374 | 38000.0 | ±9000.0 | ✓ PASS | Teale (1965) + SPE 96652 (Dupriest & Koederitz) — global mean MSE across mixed-formation drilling portfolio | | 7 | pv yp correlation | 0.7833 | 0.78 | ±0.15 | ✓ PASS | API RP-13B-1 + Bourgoyne et al. (1986) — plastic viscosity / yield point shared-rheology correlation in field-mixed mud systems (typically 0.70-0.85) | | 8 | pdc bit share | 0.7384 | 0.72 | ±0.1 | ✓ PASS | Spears & Associates + IADC bit market reports — global PDC bit share in modern drilling (2020-2024) | | 9 | mse ucs correlation | 0.7453 | 0.65 | ±0.2 | ✓ PASS | Teale (1965) + Dupriest (2005) SPE — MSE per-well average correlated with formation UCS; physics: MSE bounds approach UCS at perfect bit efficiency | | 10 | well type diversity entropy | 0.8279 | 0.85 | ±0.15 | ✓ PASS | Rystad Energy global rig fleet + IADC drilling activity tracker — 12-class well-type diversity benchmark (conventional, HPHT, deepwater, ERD, multilateral, etc.), normalized Shannon entropy | **Overall: 100.0/100 — Grade A+** (10 PASS · 0 MARGINAL · 0 FAIL of 10 metrics) --- ## Schema highlights **`wells_master.csv`** — one row per well, the relational spine. Key columns: `well_id`, `well_type` (12-class: vertical_conventional, directional, horizontal_shale, extended_reach, deepwater_offshore, hpht, geothermal, managed_pressure, underbalanced, salt_section, multilateral, slim_hole), `basin` (11-class: Permian, Eagle Ford, Bakken, Marcellus-Utica, Gulf of Mexico, North Sea, Middle East, Brazil Pre-Salt, Canada Oil Sands, Geothermal Basins, Other), `total_depth_ft`, `well_trajectory_type`, `mud_system` (water/oil/synthetic/cesium formate). **MSE follows the Teale (1965) formulation** with bit-size-aware area: > MSE = (WOB / A_bit) + (120·π·RPM·T_bit) / (A_bit · ROP) where `A_bit` varies by hole section (17.5" surface, 12.25" intermediate, 8.5" production, 6.125" lateral), and `T_bit` is downhole bit torque estimated as 25-45% of surface torque depending on trajectory and depth — a critical detail for accurate MSE in long-reach and horizontal wells. **Formation strength profiles** (`mse_log.formation_strength_psi`) follow basin-specific UCS bands with stochastic formation transitions every 500-2000 ft: Permian 8-22 kpsi, Marcellus-Utica 11-25 kpsi, Brazil Pre-Salt 15-35 kpsi, Geothermal Basins 18-40 kpsi. **Rheology coupling** — PV and YP share an underlying rheology level correlated with mud weight, producing the field-realistic PV-YP correlation of ~0.70-0.85 (Bourgoyne et al. 1986). **ECD margin** (`ecd_ppg - mud_weight_ppg`) is calibrated to maintain the ~0.45 ppg overpressure circulation envelope per IADC guidance, with depth-scaled annular friction and ROP-dependent cuttings loading. **Bit wear** uses the **IADC dull grading taxonomy** (8 codes spanning worn-teeth, broken-teeth, lost-teeth, ring-out, balled-up patterns), with dull grade escalation driven by cumulative footage and average formation strength per bit run. --- ## Suggested use cases 1. **ROP optimization regression** — predict ROP from WOB, RPM, torque, mud weight, formation UCS using the 109,870-row time-series spine. 2. **MSE efficiency classification** — train models on `bit_efficiency` to identify low-energy-efficient drilling sections. 3. **Dysfunction detection** — multi-class classifier on `dysfunction_class` (11-class: stick-slip, lateral vibration, stuck pipe, washout, twist-off, mud loss/gain, bit balling, pack-off, whirl, axial bounce) from vibration + mechanical telemetry. 4. **Stuck pipe early warning** — binary classification with the `stuck_pipe_precursor` events as positive labels, ROP/WOB/torque time- series as features. 5. **Bit dull grade prediction** — regress IADC dull grade from cumulative footage and formation strength exposure per bit run. 6. **Drilling efficiency grading** — multi-class classification on the `drilling_efficiency_grade` (A+/A/B/C/D) target from per-well aggregated features. 7. **Time-series forecasting** — predict next-stand ROP / MSE / vibration from the prior stand's drilling parameters (sequence models, transformers). 8. **Multi-table relational ML** — entity-resolution and graph-based learning across the 9 joinable tables via `well_id`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil007-sample", data_files="drilling_timeseries.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd wells = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/wells_master.csv") ts = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/drilling_timeseries.csv") mse = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/mse_log.csv") events = pd.read_csv("hf://datasets/xpertsystems/oil007-sample/drilling_events.csv") joined = ts.merge(wells, on="well_id") ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter through per-well `SeedSequence([master_seed, well_idx])` derivation, guaranteeing schema-stable joins across runs and seed-by-seed reproducibility. 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 ML prototyping and drilling- parameter research, not for live drilling operations. A few notes: 1. **Time-series is decimated 4× in the sample.** Full product runs at the generator's native cadence (~42 samples per stand at 1 Hz / 85 ft/hr). The sample uses `--decimate 4` to keep file sizes <15 MB while preserving stand-resolution detail. For high-frequency vibration ML, use the full product (12,000 wells, undecimated). 2. **ECD margin runs slightly below the 0.45 ppg target** (observed mean ~0.37 ppg) at sample scale, well within the ±0.15 ppg tolerance. This reflects the stochastic depth-scaled annular friction modeling; full- product scale converges closer to 0.45 ppg as the basin/well-type mix averages out. 3. **Max depth capped at 12,000 ft in the sample** (vs 22,000 ft in the full product). This caps the Pre-Salt and HPHT extreme-depth physics slightly under-represented. Wells assigned to those types still get their depth/mud-weight modifiers applied within the 12 kft envelope. 4. **Inter-stand jitter uses per-stand-seeded RNG** for intra-stand sample variation. This produces stand-coherent telemetry (good for ML) but means within-stand noise is correlated stand-to-stand at fixed offsets. For pure-noise modeling, filter out the per-stand structure first. 5. **Per-well stuck pipe / washout / dysfunction event rates are per-stand Bernoulli probabilities**, so total event counts per well are Poisson-distributed (~6 events/well average at sample scale). Production runs (longer wells, more stands) will see proportionally more events per well. --- ## Full product The **full OIL-007 dataset** ships at **12,000 wells**, **22,000 ft max depth**, undecimated 1Hz timeseries, full 12-class well-type coverage with HPHT and Pre-Salt extreme-depth physics, and full per-stand dysfunction event modeling — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **[e-mail]** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil007_sample_2026, title = {OIL-007: Synthetic Drilling Parameters Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil007-sample} } ``` ## Generation details - Generator version : 1.0.0 - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-21 22:54:12 UTC - Wells : 100 - Max depth : 12,000 ft (capped for sample; full product: 22,000 ft) - Timeseries decim. : 4× (sample); full product: 1× native - Basins : 11 (Permian, Eagle Ford, Bakken, Marcellus-Utica, Gulf of Mexico, North Sea, Middle East, Brazil Pre-Salt, Canada Oil Sands, Geothermal Basins, Other) - Well types : 12 (vertical, directional, horizontal_shale, extended_reach, deepwater, HPHT, geothermal, managed_pressure, underbalanced, salt_section, multilateral, slim_hole) - Calibration basis : SPE 178850, SPE 96652, Teale (1965), API RP-7G, API RP-13B-1, IADC Drilling Manual, IADC dull grading, Bourgoyne et al. (1986), Rystad Energy, Spears & Associates - Overall validation: 100.0/100 — Grade A+OIL-007 — Synthetic Drilling Parameters Dataset (Sample) is a free, schema-identical preview of XpertSystems.ais enterprise drilling-parameters dataset for ROP optimization, MSE analysis, dysfunction detection, and drilling-process ML. The sample covers 100 wells across 11 global basins and 12 well types with 220,871 rows of high-cadence drilling telemetry linked across 9 tables. It includes well master data, drilling time series, mud properties, hydraulics logs, vibration spectra, MSE logs, bit wear logs, drilling events, and drilling labels, suitable for use cases such as regression, classification, time-series forecasting, and multi-table relational ML. The data is calibrated to industry standards and reproducible, though the sample has limitations in time-series decimation, depth capping, and scale.

搜集汇总

数据集介绍

构建方式

在油气上游钻井作业的数字化与智能化转型背景下，该数据集的构建立足于真实钻井工程物理规律与行业标准的多重约束。生成器以确定性随机种子驱动，通过逐井派生种子序列，对全球11个盆地、12类井型进行层次化采样，并沿井深随机模拟地层强度跃迁，每500至2000英尺切换一次UCS区间。机械钻速、钻压、扭矩、转速等高频时序参数由钻头尺寸、轨迹类型和钻井液流变特性耦合生成，其中机械比能严格采用Teale（1965）公式，并考虑井下钻头扭矩占地面扭矩25%至45%的衰减。泥浆性能则通过共享流变层级确保塑性黏度与动切力呈现0.70至0.85的现场相关性。这种“物理模型引导+随机过程驱动”的构建范式，使样本在统计特征与工程逻辑上逼近真实钻井场景。

使用方法

该数据集面向钻井过程机器学习与钻速优化研究，支持多类典型任务。用户可通过Hugging Face datasets库或pandas直接加载CSV文件，并利用井标识符进行多表关联，构建关系型机器学习管道。具体应用涵盖：以钻压、转速、扭矩、泥浆密度和地层强度为特征，对机械钻速进行回归预测；基于振动与机械遥测信号，对卡钻、涡动、粘滑等11类功能障碍进行多分类检测；利用累积进尺与地层暴露强度预测IADC钻头钝化等级；或针对单井聚合特征进行钻井效率分级。时序建模者可借助前一立柱的钻井参数序列，预测下一立柱的钻速、机械比能及振动状态，亦可开展多表图学习与实体解析。所有生成过程保持确定性随机种子，确保结果可复现。

背景与挑战

背景概述

石油天然气上游钻井作业中，机械钻速优化与钻井效率分析长期依赖现场经验与事后数据，制约了机器学习在该领域的深入应用。2026年，XpertSystems.ai发布了OIL-007合成钻井参数数据集，旨在为钻速优化、机械比能分析及钻井异常检测提供高保真、多表关联的基准数据。该样本数据集涵盖全球11个盆地、12种井型、100口井的22万余行高频遥测记录，严格依据SPE、API及IADC等行业标准进行校准，其10项指标验证均达到A+等级，为钻井过程建模与算法评测建立了可复现的行业标杆。

当前挑战

该数据集所面临的挑战具有双重维度。在领域问题层面，钻井过程的多变量耦合、地层不确定性及井下异常事件的稀疏性，使得钻速预测与 dysfunction 分类远较常规时间序列任务复杂，且真实数据因商业保密而极度稀缺。在构建过程中，合成数据需在物理合理性与统计真实性之间取得平衡：既要忠实还原机械比能随井深与地层强度的变化规律，又要保证泥浆流变参数、钻头磨损与振动频谱之间的跨表一致性；同时，样本在时间序列抽稀、最大井深截断及事件泊松分布等方面的局限性，也对模型泛化能力提出了额外考验。

常用场景

经典使用场景

在油气上游钻井工程领域，机械钻速优化与井下异常工况识别长期依赖现场经验与物理模型，数据驱动的建模范式则亟需高保真、多表关联的钻井遥测语料作为支撑。OIL-007样本数据集以109,870行高密度时间序列为主干，涵盖钻压、转速、扭矩、立管压力与悬重等关键参数，并借助井级主表实现九表关联，为ROP回归预测、MSE效率分类、黏滑与井漏等十一类 dysfunction 检测以及卡钻早期预警等经典任务提供了同一化建模基础，尤其适配序列模型与多表关系学习的实验需求。

解决学术问题

钻井过程建模研究长期面临真实数据获取成本高昂、跨盆地异质性显著以及标签体系缺失等瓶颈，致使模型泛化能力难以验证。该数据集以Teale机械比能公式为物理内核，融合IADC钻头磨损分级与API流变学标准，预设最优ROP、 dysfunction 风险与钻井效率等级等六类机器学习标签，使研究者能够在可控生成机制下探讨RO P主控因素、比能—地层强度耦合关系以及多任务学习框架，为钻井参数优化与异常识别的可复现研究确立了基准。

实际应用

在工程实践中，该数据集可服务于钻井参数实时优化、钻头磨损趋势预判与井下复杂情况告警等场景。工程师可基于振动频谱与机械遥测构建黏滑、涡动与轴向跳动的多分类器，也可利用卡钻前兆标签训练早期预警模型，从而在钻进过程中动态调整钻压与转速组合，降低非生产时间。其井型覆盖直井、水平井、高温高压井与深水井等十二类，盆地横跨二叠纪、北海与巴西盐下等十一区，为跨工况迁移验证提供了贴近现场的应用环境。

数据集最近研究

最新研究方向

围绕钻井参数优化与井下复杂工况智能预警，oIL007-sample正推动数据驱动范式从单表建模向多表关系学习与物理约束融合演进。研究者借助该数据集高保真的机械比能与振动频谱特征，探索基于Transformer的钻速序列预测及黏滑、井漏等11类 dysfunction 事件的早期辨识；同时，将Teale模型与IADC磨损分级嵌入损失函数，构建物理信息神经网络以提升跨盆地泛化能力。此类工作不仅为自动化钻井闭环控制提供了可复现的基准，亦回应了油气上游数字化转型中对可解释、可迁移机器学习模型的迫切需求，对降低非生产时间与钻完井成本具有显著意义。