NoraResearchLab/lithology-sequence-benchmark

Lithology Sequence Identification Benchmark

Evaluation-only benchmark: predict the lithological sequence of a whole well from raw wireline logs

---

Task

Given one complete well log (curves sampled every 0.5 ft), output the ordered list of continuous lithological intervals across depth:

[{"top_depth_ft": 2150.0, "base_depth_ft": 2191.5, "lithology": "SHALE"},
 {"top_depth_ft": 2191.5, "base_depth_ft": 2230.0, "lithology": "LIMESTONE"},
 {"top_depth_ft": 2230.0, "base_depth_ft": 2246.5, "lithology": "UNKNOWN/MIXED"}]

Intervals must be contiguous and cover the logged interval. Classes: SHALE, SANDSTONE, LIMESTONE, DOLOMITE, ANHYDRITE, SALT, COAL, UNKNOWN/MIXED. UNKNOWN/MIXED marks depths where the available logs do not support a reliable single-lithology call.

Labels are withheld. The public release contains only the QC'd logs and well metadata. Reference intervals are retained privately by the maintainers for offline scoring of submitted predictions.

Wells	245 (train 166 / validation 25 / test 54 — split by well)
Depth samples	1,963,283
Reference intervals	49,830
Label-quality tiers	A: 167, B: 78
Config fingerprint	7b440d013c0a08a1
Reference fingerprint	92e77d413b4c86d4

Files / configs

Config	Content	Role
logs	well_id, depth_ft, GR, RHOB, NPHI, PEF, DT, RT, CALI (QC'd, NaN = curve absent)	model input
logs	well_id, depth_ft, GR, RHOB, NPHI, PEF, DT, RT, CALI (QC'd, NaN = curve absent)	model input
wells	one row per well: split, tier, curves present, baselines, interval count	metadata
intervals / depthwise	withheld – private reference held by maintainers	evaluation only

label_config.json holds every threshold used; curve_specs.json the mnemonic rules; manifest.json versions and hashes.

How the reference labels are produced (deterministic, no randomness, no machine learning)

1. Curve detection — each LAS curve is matched to a standard curve by mnemonic alias → regex → description keywords → unit, so RHOZ, DEN, RHOB all become RHOB, ILD/LLD/AT90/conductivity become RT, etc. Ties are broken by data coverage then name.
2. Unit standardisation — metres→feet, kg/m³→g/cc, %→fraction, µs/m→µs/ft, mm→in, conductivity→resistivity.
3. Quality control — sentinel/null masking, physical-range masking, flat-line (dead tool) masking, Hampel spike removal, resampling to a fixed grid (bin-mean for fine logs, gap-limited interpolation for coarse), low-coverage curve rejection, caliper-based washout flag that down-weights density/neutron/PEF/sonic evidence.
4. Multi-log fuzzy scoring — per depth, each lithology is scored from every available feature (gamma index normalised per well, RHOB, limestone-scale NPHI, PEF, DT, log resistivity, neutron–density separation) through trapezoidal membership functions and configurable weights. Missing curves reduce evidence; they never crash the method. Contradicted lithologies are penalised.
5. UNKNOWN/MIXED gate — assigned when available-evidence weight, best score, or the margin over the runner-up falls below configured limits.
6. Segmentation — depth-wise calls are median/mean smoothed, then runs thinner than the minimum bed thickness are iteratively absorbed into the better-fitting neighbour and equal neighbours merged, yielding geological intervals instead of log noise.

Lithology plateau ranges used (full trapezoids in label_config.json):

Lithology	GRI	RHOB	NPHI	PEF	DT	LOGRT	ND_SEP
SHALE	0.65 – 2.5	2.2 – 2.65	0.25 – 0.45	2.5 – 3.8	85 – 140	-0.2 – 0.7	0.05 – 0.22
SANDSTONE	-1 – 0.25	2.15 – 2.65	-0.04 – 0.18	1.6 – 2.3	52 – 90	0 – 2.2	-0.12 – -0.02
LIMESTONE	-1 – 0.25	2.35 – 2.72	-0.01 – 0.18	4.7 – 5.5	46 – 72	0.9 – 3	-0.02 – 0.03
DOLOMITE	-1 – 0.28	2.65 – 2.9	0.02 – 0.2	2.9 – 3.5	43 – 65	0.8 – 3	0.06 – 0.16
ANHYDRITE	-1 – 0.12	2.92 – 3	-0.04 – 0.03	4.8 – 5.4	48 – 53	2.6 – 5	0.12 – 0.22
SALT	-1 – 0.12	2 – 2.15	-0.04 – 0.04	4.4 – 4.9	65 – 70	2.6 – 5	-0.45 – -0.25
COAL	-1 – 0.55	1.2 – 1.7	0.45 – 0.85	0.2 – 1.5	110 – 160	1.3 – 3.5	-0.25 – 0.25

Evaluation

from datasets import load_dataset
import pandas as pd
logs = pd.concat([load_dataset("NoraResearchLab/Lithology-Sequence-Benchmark", "logs", split=s).to_pandas() for s in ["test"]])
ref  = pd.concat([load_dataset("NoraResearchLab/Lithology-Sequence-Benchmark", "intervals", split=s).to_pandas() for s in ["test"]])

# my_model(well_curves_df) -> DataFrame[top_depth_ft, base_depth_ft, lithology]
# summary, per_well = evaluate_model(my_model, logs, ref)      # evaluate_model / metrics live in the builder notebook

Reported metrics: pooled depth macro-F1 / accuracy (also on confident-reference depths only), boundary F1 at ±2.5/5/10 ft, and sequence similarity (1 − normalised edit distance of the ordered lithology sequence).

Important limitations — please read

• These are rule-derived (silver-standard) labels, not core- or cutting-verified lithology. They measure agreement with a transparent petrophysical interpretation of the logs; they are not independent ground truth. A model with access to the same curves can in principle learn the rules, so strong scores indicate rule recovery, not proven geological accuracy.
• Default signatures are generic clean-matrix / textbook ranges, not calibrated to Kansas. Boundaries shift by a few feet because of smoothing.
• Neutron logs are assumed limestone-scale unless the mnemonic says otherwise. Heavy mud, casing, salt-section washouts and bad tools can mislead any log-only method.
• Sandstone-vs-limestone-vs-dolomite separation is weaker when PEF/DT are absent (tier B wells), which increases UNKNOWN/MIXED.

Source data & attribution

Logs come from the public digital wireline-log archives of the Kansas Geological Survey (KGS), University of Kansas (https://www.kgs.ku.edu/Magellan/Logs/). KGS distributes data "as is" without warranty; consult the KGS terms and disclaimer before downstream use. This is an independent derivative work, not endorsed by the KGS.

Maintainer

NORA Research Lab

Quick links: Website · GitHub · Hugging Face · LinkedIn · X