CrudeOilMix: A Million-Scale Multimodal Benchmark for Crude Oil Characterization

Paper: NeurIPS 2026 Evaluations & Datasets Track (under review)

Dataset Summary

CrudeOilMix is a multimodal benchmark comprising 1,141,933 crude oil samples generated by blending 9,061 real crude oil assays through H/CAMS, an industry-standard refinery simulation platform.

Each sample provides three aligned modalities:

• wc_ent — whole-crude entered (laboratory-equivalent) properties: 187 attributes, highly sparse
• wc_calc — whole-crude calculated (H/CAMS-derived) properties: 187 attributes, 64 fully observed
• cuts — nine distillation-cut fractions, each sharing the same 187-attribute schema
• tbp — four true-boiling-point curves (up to 41 points each)
• info — blend recipe and crude metadata

The dataset mirrors real-world refinery operations:

• 2-crude blends: 407,031 samples
• 3-crude blends: 510,977 samples
• 4-crude blends: 214,864 samples

Benchmark Tasks

Task 1 — Blend Property Prediction: Given properties of up to 4 component crudes and their mixing ratios, predict 15 whole-crude properties of the resulting blend.

Task 2 — Property Imputation: Given a partially observed property vector (30–70% masked), predict values at masked positions.

Two split protocols are provided:

• Split-R (global): Random 80/10/10 split (train/val/test)
• Split-C (composition): Held-out crude combinations to test generalization to unseen blends

Repository Structure

data/                         Full dataset (14 GB)
  wc_ent_part00000.parquet    Whole-crude entered properties (58 parts)
  wc_calc_part00000.parquet   Whole-crude calculated properties (58 parts)
  cuts_part00000.parquet      Distillation-cut fractions (58 parts)
  tbp_part00000.parquet       True-boiling-point curves (58 parts)
  info_part00000.parquet      Crude metadata (58 parts)
  blend_mix_manifest.parquet  Blend recipes and n_components for all samples
  split_global.parquet        Split-R assignments (train/val/test)
  split_c_components.parquet  Split-C crude-level assignments
  split_c_task1_blends.parquet Split-C blend-level assignments

sample/                       Representative sample (10.9 MB, 800 samples)
  *_sample.parquet            One file per modality, 200 per blend complexity

Loading the Data

import pandas as pd

# Load the full wc_calc modality
import glob
parts = sorted(glob.glob("data/wc_calc_part*.parquet"))
wc_calc = pd.concat([pd.read_parquet(p) for p in parts], ignore_index=True)

# Or load the small sample (800 rows, ~11 MB total)
sample = pd.read_parquet("sample/wc_calc_sample.parquet")

# Load blend recipes
manifest = pd.read_parquet("data/blend_mix_manifest.parquet")
# Columns: oilid, n_components, is_blend, mix_json

# Load train/val/test splits
splits = pd.read_parquet("data/split_global.parquet")
train_ids = splits[splits.split == "train"]["oilid"]

Physical Consistency

The dataset satisfies six petroleum-science invariants verified across all 1,141,933 samples:

• ASTM API–SG conversion: zero error (API = 141.5/SG − 131.5)
• Distillation temperature ordering: T10 < T50 < T90 (100% compliance)
• Mass balance conservation across distillation cuts
• Viscosity–temperature monotonicity (SV15 ≤ SV20)
• Density ordering (DN15 ≤ DN20)
• Yield–cut ordering

Citation

@inproceedings{crudeolimix2026,
  title     = {{CrudeOilMix}: A Million-Scale Multimodal Benchmark for Crude Oil Blend Evaluation},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
}

License

CC BY 4.0

The dataset is generated from H/CAMS industrial simulation software using anonymized crude oil assay data. It contains no personally identifiable information and no proprietary assay values.