--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - downstream - catalyst - deactivation - arrhenius - fcc - hydrocracker - predictive-maintenance - xpertsystems pretty_name: "OIL-023 — Synthetic Catalyst Degradation Dataset (Sample)" size_categories: - 100K8 wt% indicates accelerated deactivation requiring regeneration) | | 5 | avg pressure drop psi | 19.1523 | 19.0 | ±6.0 | ✓ PASS | Ergun equation (1952) packed bed pressure drop + UOP design spec — typical operating pressure drop for mixed FCC/hydroprocessing reactor portfolio (8-25 psi typical; >2x design indicates fouling) | | 6 | avg regeneration efficiency | 0.8817 | 0.88 | ±0.06 | ✓ PASS | Bartholomew (2001) regeneration kinetics + UOP/BASF FCC regenerator data — mean cycle regeneration efficiency for properly-managed catalyst (85-95% typical; declines with cycle count due to thermal damage) | | 7 | activity coke pearson correlation | -0.9780 | -0.85 | ±0.15 | ✓ PASS | Bartholomew (2001) + Arrhenius (1889) — expected strong inverse correlation between catalyst activity and coke loading (coupled exponential decay: as activity declines via base_decay = exp(-severity*age/life), coke accumulates as 1 - base_decay). Validates generator's Arrhenius-style deactivation physics. | | 8 | activity conversion pearson correlation | 0.6917 | 0.6 | ±0.15 | ✓ PASS | Levenspiel chemical reactor engineering + Bartholomew (2001) — expected strong positive correlation between catalyst activity and conversion percentage (conversion ∝ activity per first-order kinetics with Sabatier-style poison terms). Validates generator's kinetics coupling. | | 9 | health shutdown risk pearson correlation | -0.9982 | -0.95 | ±0.1 | ✓ PASS | Generator's deterministic formula: shutdown_risk = (100 - health_score)/100 + anomaly*0.18. Expected near-perfect inverse coupling. Validates feature-coupled label generation for predictive maintenance ML applicability. | | 10 | process unit diversity entropy | 0.9793 | 0.92 | ±0.05 | ✓ PASS | 11-class process unit taxonomy per UOP/Honeywell + Axens refinery licensing portfolio (FCC, Hydrocracker, Hydrotreater, Catalytic Reformer, Isomerization, Alkylation, Resid Hydroprocessing, Renewable Diesel HT, Steam Methane Reformer, Sulfur Recovery, Aromatics Unit), normalized Shannon entropy | **Overall: 100.0/100 — Grade A+** (10 PASS · 0 MARGINAL · 0 FAIL of 10 metrics) --- ## Schema highlights **`reactors_master.csv`** — 11-class process unit taxonomy with unit-specific design specs: | Process Unit | Design T (°F) | Design P (psi) | Conversion (%) | Catalyst Life (days) | |---|---:|---:|---:|---:| | FCC | 960 | 32 | 76 | 90 | | Hydrocracker | 760 | 1900 | 82 | 720 | | Hydrotreater | 690 | 1100 | 91 | 900 | | Catalytic Reformer | 940 | 180 | 86 | 540 | | Isomerization | 330 | 420 | 83 | 730 | | Alkylation | 160 | 180 | 88 | 365 | | Resid Hydroprocessing | 780 | 2300 | 74 | 540 | | Renewable Diesel HT | 660 | 1300 | 93 | 640 | | Steam Methane Reformer | 1550 | 350 | 89 | 1460 | | Sulfur Recovery Unit | 640 | 12 | 96 | 1095 | | Aromatics Unit | 880 | 230 | 84 | 600 | **`catalyst_activity.csv`** — **Bartholomew (2001) exponential decay** implementation: > base_decay = exp(-severity × age_days / nominal_life) > activity = 100 × (0.22 + 0.78 × base_decay) + regen_boost Activity declines from ~100% fresh to ~22% fully-deactivated, with severity (0.55-1.65) modulating decay rate and regeneration cycles partially recovering activity (7-18 percentage point boost per cycle, declining with thermal damage). **`coke_deposition.csv`** — **Bartholomew (2001) coupled coke accumulation**: > coke = 0.6 + 9.8 × (1 - base_decay) × severity − 0.055 × regen_boost Coke loading rises from <1 wt% fresh to 10+ wt% near deactivation. The sample's activity↔coke Pearson correlation is r ≈ **−0.98** — **near- deterministic inverse coupling per Arrhenius physics**. **`pressure_drop_profiles.csv`** — **Ergun (1952) packed-bed pressure drop**: > pressure_factor = 1 + coke/12 + severity × age/(3.2 × nominal_life) > pressure_drop = design_dp × pressure_factor + noise **`conversion_efficiency.csv`** — kinetics-coupled conversion + selectivity: > conversion = target × (0.78 + 0.22 × activity/100) - 0.035 × sulfur/10 + noise > selectivity = fresh × (0.86 + 0.14 × activity/100) - 0.06 × coke + noise > yield = conversion × selectivity/100 The sample's activity↔conversion Pearson correlation is r ≈ **+0.69** — **strong positive coupling per first-order reactor kinetics**. **`catalyst_labels.csv`** — **deterministic feature-coupled labels**: > health_score = 0.48 × activity + 0.24 × (100 - coke × 4.2) > + 0.18 × (100 - dp_ratio × 16) + 10 × h2_util > replacement_priority = 'high' if health < 45 OR dp > 2.6 × design_dp > else 'medium' if health < 65 OR dp > 1.8 × design_dp > else 'low' > shutdown_risk_score = (100 - health_score)/100 + anomaly × 0.18 The sample's health↔shutdown Pearson correlation is r ≈ **−0.998** — **near-deterministic inverse coupling validates label generation formula**. --- ## Suggested use cases 1. **Catalyst RUL (Remaining Useful Life) regression** — predict `estimated_remaining_life_days` from operating features per Bartholomew deactivation kinetics. **Strong physics signal**: activity-coke r ≈ −0.98. 2. **3-class replacement priority classification** — multi-class classifier on `replacement_priority` from health features. **Strong feature coupling** — models WILL learn meaningful patterns. 3. **Activity decay regression** — predict `relative_activity_pct` from age + severity + regen history. Pure Arrhenius signal. 4. **Coke deposition forecasting** — time-series forecasting of `coke_loading_wt_pct` per coupled decay physics. 5. **Conversion-yield prediction** — predict `yield_pct` from activity + coke + sulfur features per kinetics. 6. **Regeneration cycle ROI optimization** — regression on `replacement_roi_score` from cumulative thermal damage + cycle count features. 7. **Catalyst failure root cause classification** — 12-class classifier on `root_cause` (rare events; see Honest Disclosure §3). 8. **Emissions prediction** — regression on `co2_tpd` / `nox_ppm` / `sox_ppm` from operating + coke + sulfur features per EPA NSPS Subpart Ja. 9. **Anomaly detection** — multi-variate anomaly detection on poisoning + activity + coke time series. 10. **Multi-table relational ML** — entity-resolution and graph neural-network learning across the 13 joinable tables via `reactor_id`, `catalyst_id`, `timestamp`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil023-sample", data_files="catalyst_activity.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd reactors = pd.read_csv("hf://datasets/xpertsystems/oil023-sample/reactors_master.csv") act = pd.read_csv("hf://datasets/xpertsystems/oil023-sample/catalyst_activity.csv") coke = pd.read_csv("hf://datasets/xpertsystems/oil023-sample/coke_deposition.csv") conv = pd.read_csv("hf://datasets/xpertsystems/oil023-sample/conversion_efficiency.csv") labels = pd.read_csv("hf://datasets/xpertsystems/oil023-sample/catalyst_labels.csv") # Full Arrhenius+kinetics feature engineering: joined = (act .merge(coke, on=["reactor_id", "timestamp"]) .merge(conv, on=["reactor_id", "timestamp"]) .merge(labels, on=["reactor_id", "timestamp"]) .merge(reactors, on="reactor_id")) # Predict replacement_priority from activity + coke + conversion + design specs ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving `np.random.default_rng` plus python `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 catalyst degradation ML research, not for live operational decisions. Several notes: 1. **Catalyst master has more catalysts than are actively used.** The generator creates `n_catalysts` catalyst lots in `catalyst_master.csv` but **only uses the FIRST catalyst per reactor** in the time-series simulation (via `group.iloc[0]`). With 200 catalysts and 50 reactors, **150 catalysts in master are not referenced by any time-series table.** Treat `catalyst_master.csv` as a **vendor portfolio reference** rather than fully-linked operational data. To filter to only operationally-active catalysts: ```python active_ids = set(activity['catalyst_id'].unique()) cats_active = cats[cats['catalyst_id'].isin(active_ids)] ``` 2. **Anomaly rate is per-timestep, not cumulative.** The generator's `anomaly_injection_rate=0.032` is the **annual** anomaly probability divided by 365 → ~0.0088% per daily timestep in the sample. Cumulative anomalies over 365 days approach the 3.2% annual target. **Observed per-row anomaly rate is ~0.002** in the sample — this is correct generator behavior, not a bug. For event-classification ML, aggregate to per-reactor-week or per-reactor-month windows. 3. **Failures are very sparse** (~2-5 events per 50 reactors at sample scale). Failure events require `replacement_priority == 'high' AND rng < 0.0025` OR `anomaly_flag AND rng < 0.12`, which creates rare events for ML class-balancing. For 12-class root cause classification, **use the full product** (1500+ reactors) or merge failure events from OIL-021 / OIL-022 for richer event populations. 4. **Sulfur↔conversion correlation is weak (r ≈ −0.03)** because sulfur is sampled per-timestep from `lognormal × feed_contam_bias`, so each reactor's sulfur signal is dominated by its own bias level rather than time-evolving. **For Sabatier-style poisoning ML, normalize sulfur per-reactor first** (z-score within reactor_id) before fitting models. 5. **Coke↔pressure drop correlation is moderate (r ≈ 0.11)** — weaker than expected because the pressure drop formula uses `design_dp × pressure_factor + noise` where `design_dp` varies substantially across the 11 process unit types (6-28 psi range). Pressure drop is dominated by cross-reactor unit-type variance rather than within-reactor coke evolution. **For Ergun-style pressure drop ML, normalize dp per-reactor** (dp_ratio = pressure_drop / design_pressure_drop) before fitting. 6. **Replacement priority is heavily 'low'-dominant (~89%)** at sample scale because the formula triggers `high` only at health < 45 or dp > 2.6× design — most sample timesteps have moderate degradation. The 3-class distribution becomes more balanced at longer simulation horizons (3650-day prod mode). For class-balanced 3-class classification, **oversample medium/high labels or weight loss appropriately**. 7. **Hydrogen utilization is zero for 5 of 11 process units** (FCC, Alkylation, SMR, SRU, plus most coker units have h2=0 by design). This means `hydrogen_utilization_efficiency` will be near-zero for approximately half the reactor portfolio. **Filter to hydroprocessing units** (Hydrocracker, Hydrotreater, Resid HP, Renewable Diesel HT, Catalytic Reformer, Aromatics, Isomerization) for H2-related ML. 8. **Steam Methane Reformer is hottest (1550°F) and skews reactor_operations.reactor_temp_f distribution.** The 11-unit portfolio spans 160°F (Alkylation) to 1550°F (SMR) — for temperature-feature ML, **either filter to a single unit type or one-hot encode unit type as a feature** to avoid temperature as a proxy for unit type. --- ## Cross-references to other XpertSystems OIL SKUs This SKU is the **fourth downstream (refining) SKU** in the catalog — specializing in **catalyst lifecycle physics**: | SKU | Layer | Focus | |---|---|---| | OIL-019 | Downstream — process | Refinery unit operations (CDU/VDU/FCC + control + HX) | | OIL-020 | Downstream — yield | Crude → product yields + economics + emissions | | OIL-022 | Downstream — turnaround | Turnaround planning + RBI + inspection | | **OIL-023** | **Downstream — catalyst** | **Catalyst deactivation physics + regeneration + RUL** *(this SKU)* | | OIL-021 | Cross-stream | Equipment performance + condition monitoring | **OIL-023 vs OIL-019/020/022**: OIL-019 simulates **steady-state refinery process operations** (control loops, heat exchangers). OIL-020 simulates **aggregate refinery yields + economics**. OIL-022 simulates **turnaround / shutdown / inspection events**. **OIL-023 specializes in the catalyst lifecycle itself** — the continuous-time degradation physics that drives turnaround timing decisions in OIL-022. Use OIL-023 for **catalyst ML and predictive maintenance**, OIL-022 for **turnaround planning ML**. **OIL-023 vs OIL-021**: OIL-021 simulates **rotating + static equipment performance** (HX, compressors, pumps, motors). OIL-023 specializes in **catalyst-bearing reactor performance** (FCC, hydrocracker, hydrotreater). Use OIL-021 for rotating-equipment PHM, OIL-023 for catalyst PHM. --- ## Full product The **full OIL-023 dataset** ships at **1,500 reactors × 3,650 days × 12,000 catalyst lots** (prod mode) producing tens of millions of rows with **clearer class-balanced replacement priority distributions** (long-horizon simulation drives more high-priority transitions), **richer failure event populations** (200+ failures per 1,500 reactors for class-balanced 12-class root cause ML), and **stronger sulfur-conversion coupling** at scale — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil023_sample_2026, title = {OIL-023: Synthetic Catalyst Degradation Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil023-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-22 20:55:20 UTC - Reactors : 50 - Catalyst lots : 200 (in master; 50 actively used in time-series) - Simulation days : 365 - Time-step freq : 24 hours (daily) - Process units : 11 (FCC, Hydrocracker, Hydrotreater, Catalytic Reformer, Isomerization, Alkylation, Resid Hydroprocessing, Renewable Diesel HT, Steam Methane Reformer, Sulfur Recovery Unit, Aromatics Unit) - Catalyst types : 22 (zeolite Y, ZSM-5, NiMo / CoMo alumina, Pt-Re / Pt-Sn alumina, sulfided NiMo, noble-metal HDO, nickel alumina/magnesia, titania Claus, etc.) - Vendors : 11 (Albemarle, BASF, UOP/Honeywell, Axens, Topsoe, Johnson Matthey, Criterion, Shell Catalysts, Clariant, W.R. Grace, Sinopec Catalyst) - Failure root causes: 12 (coke runaway, sulfur poisoning, nitrogen poisoning, metals fouling, thermal sintering, bed channeling, pressure drop excursion, feed contamination, oxygen breakthrough during regen, steam aging, mechanical attrition, chloride imbalance) - Regions : 10 - Calibration basis : Bartholomew (2001), Forzatti & Lietti (1999), Arrhenius (1889), Ergun (1952), API RP 939-C, NACE TM0185, ASTM D5757/D7964/D4567/D5630, UOP/Topsoe/Albemarle/BASF, EPA NSPS Subpart Ja, Levenspiel reactor engineering - Overall validation: 100.0/100 — Grade A+