Horizontal Well Log Datasets for Geosteering: What You Can Actually Train On

Nothing else in the industry's 14 primary cataloged datasets pairs horizontal trajectories with gamma-ray curves and interpreted targets. OGIM v2.5.1 maps about 6.7 million surface features but records no subsurface curves; NORA 3D is LiDAR vision data; the Multimodal RAG corpus serves PDF retrieval. Datadory's brief calls ROGII's release the largest openly shared geosteering training set in the slice, which makes it the anchor for any TVT-prediction project.

The horizontal well CSV arrives with a fixed column list captured during research: WELLNAME, MD, X, Y, Z, ANCC, ASTNU, ASTNL, EGFDU, EGFDL, BUDA, TVT, GR, TVT_input. Wells are keyed by an 8-character hash such as 015fe0d2, shared between a well's horizontal file and its typewell so joins stay trivial.

Each column has a job:

1. MD - measured depth in feet, the total wellbore length from surface.
2. X / Y - easting and northing coordinates of each wellbore point, in feet.
3. Z - true vertical depth below sea level.
4. GR - gamma ray in API units, present in both the horizontal file and the typewell; this is the curve your model correlates on.
5. ANCC, ASTNU, ASTNL, EGFDU, EGFDL, BUDA - predicted depths of six named formation horizons (training files only).
6. TVT - true vertical thickness in feet, the manually interpreted geological position per foot of lateral, and the competition target.
7. TVT_input - a copy of TVT handed to you as an input feature, with NaNs masking the evaluation zone.

That masked zone matters more than any single column: local validation must come from train/ splits because the visible test folder holds only a few authoring examples before the hidden test set is substituted at rerun.

A typewell is the vertical reference log for the area - in this corpus a three-column CSV (TVT, GR, Geology) pairing the thickness index with the same gamma-ray signature used downhole, plus a categorical formation label such as EGFDL attached to rows for correlation. Geosteering interpretation works by sliding the measured gamma-ray curve against this reference until the match locks in; every manual TVD pick a geologist draws on a cross-section is essentially that correlation done by eye.

For an ML engineer the typewell converts the problem from sequence regression into something closer to alignment: the model learns a mapping from the lateral's GR signature onto the typewell's GR-versus-TVT profile. The submission format reflects how granular that gets - one row per foot per well, keyed {WELLNAME}_{row_index} (example 015fe0d2_1654) with the predicted tvt value, scored by RMSE. Sample rows show values like -38.7 ft, small offsets that still decide whether a lateral stays in zone over thousands of feet.

When you need full curves rather than survey tables, the NSTA National Data Repository holds roughly 1.6 PB of SEG-Y, LAS and DLIS files, royalty-free after free registration subject to attribution and monthly company quotas. Pulling LAS from real offset wells gives you additional gamma-ray character to sanity-check how well a model trained on the competition basin generalizes elsewhere.

A working pipeline fits inside a notebook session, which is exactly how the competition scored submissions:

Static releases like this never grow - only 181 of the 1,744 datasets Datadory catalogs share the static label - so treat the corpus as a benchmark, not a live feed.

Head-to-head comparisons inside the slice make the boundaries explicit: see vs OPD - Offshore Platform Dataset (Sentinel-1 derived) for the subsurface-versus-surface trade quantified.