Industries / Energy, water and waste / Oil and gas extraction
Exploration and seismic surveys
Exploration is the search for places where oil and gas are trapped underground, mostly by sending sound into the earth and reading the echoes, and by measuring the rock inside drilled holes.
6
Milestones achieved, from 1921 to 2018.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1921
Echoes map the rock
On 9 August 1921 a team in Oklahoma measured the world's first reflection seismograph geologic section. The method sets off a small charge, records the sound bouncing back from rock layers, and works out the depth of each layer from the timing. On 4 December 1928 Amerada struck oil near Seminole, Oklahoma, in a structure found this way, which the American Oil & Gas Historical Society calls the world's first such discovery.
Credit
John Karcher, William Haseman, David Ohern and Irving Perrine; Amerada Petroleum Corporation
1927
Reading rock inside a well
On 5 September 1927 a Schlumberger team led by Henri Doll lowered an electrical probe into a 500-meter well at Pechelbronn, France, and plotted the rock's resistivity (how strongly it resists an electric current) against depth. Oil-bearing rock resists current more than rock full of salt water, so the hand-plotted graph showed where to look. SLB calls it the first electrical resistivity well log.
Credit
Conrad and Marcel Schlumberger, Henri Doll
1967
Seismic in three dimensions
Exxon shot the first 3D seismic survey at its Friendswood field near Houston in 1967, according to a 2005 Oil & Gas Financial Journal article. A 3D survey records echoes on a grid instead of along one line, so it gives a block image of the rock instead of a slice. Within five years Chevron, Texaco, Mobil, Phillips, Amoco and Unocal had 3D projects, the same article says.
Credit
Exxon
1983
Mukluk, the costly dry hole
Sohio, majority-owned by BP, began drilling Mukluk No. 1 in Alaska's Beaufort Sea in November 1983 from an island built in 20 meters of water. The reservoir rock was good but held water, not oil. Geologist David Bamford called it "the world's most expensive dry hole, at least up to that point."
Lessona seismic image shows the shape of a trap, not whether oil is still in it, so doubts about when the trap formed or whether it leaked are real risks, not loose ends.
Credit
Sohio
2003
Watching a field change
In summer 2003, 120 km of seismic cable with about 2,400 receivers was trenched into the seafloor over 45 km² of Norway's Valhall field. The field could then be surveyed again and again to see where injected water was moving, a method called time-lapse or 4D seismic. The same article dates the first permanent seafloor array to the Foinaven project in UK waters in 1995.
Credit
BP and the Valhall partners
2018
A whole field, made public
Equinor and its partners released about 40,000 files holding all subsurface and operating data from the Volve field in the North Sea, which produced from 2008 to 2016. The data is free for research, study and development under the Equinor Open Data Licence. Students now have a complete real field to practise on.
Credit
Equinor and the Volve licence partners
Watch for
Seismic AI model goes commercial
TGS says its Seismic Foundation Model, trained on TGS's global library of seismic surveys, will support structural and stratigraphic interpretation for an unnamed international oil company under a multi-year contract. A foundation model is a large AI model trained on raw data first and then adapted to many tasks. TGS gives no measured results and names no outside evaluator.
Achieved whenan outside party publishes the model's accuracy on public labelled seismic data, or the client reports measured gains.
Claimed by
TGS, August 2026
Open seismic AI model
The Norwegian Computing Center (NR), working with Aker BP and Equinor, released the NCS models, trained on public seismic data from Norway's DISKOS repository, with weights on Hugging Face and code on GitHub. Its paper reports that the NCS-2.5D model had the highest average score of the models it tested.
Achieved whena group outside NR and its partners reproduces the scores on benchmarks of its own choosing.
Claimed by
NR, March 2026
Open
01
Read old paper well logs
A tool that turns a scanned paper well log (a long strip chart of rock measurements against depth) into a digital curve file in the standard LAS format, and states its error.
The briefCredit · unclaimed02
Test the seismic AI models
An independent, repeatable test of the public NCS seismic models, and of any other seismic model whose weights are released, on public labelled data their developers did not train on.
The briefCredit · unclaimed
Open 01 · the brief
Read old paper well logs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that turns a scanned paper well log (a long strip chart of rock measurements against depth) into a digital curve file in the standard LAS format, and states its error.
Why it is still open
The best published deep-learning attempt, VeerNet (2022), picked out curves with an F1 score of 35% and an overlap score (IoU) of 30%, and its gamma-ray values correlated with the true values at only 0.62. Its authors note that the usual alternative is a costly digitizer that is "manual and time-consuming." The raw material is public: Colorado's regulator says it has scanned "a fairly complete set" of the state's well logs.
Built with what exists
Vision-language models that read images; open-weight models that can run on your own hardware (Note: "A trillion open parameters"); the VeerNet method as a published baseline; scanned logs from state archives.
First step
Find 20 wells in one public archive that have both a scanned log and an LAS file, and score a baseline method on them.
Done looks like
On at least 100 public wells that have both a scanned log and a digital LAS file, the digitised gamma-ray and resistivity curves correlate with the LAS values at 0.95 or better, with the tool, the well list and the scores published.
Open 02 · the brief
Test the seismic AI models
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent, repeatable test of the public NCS seismic models, and of any other seismic model whose weights are released, on public labelled data their developers did not train on.
Why it is still open
NR's page reports no outside evaluation of its models, and TGS gives no figures for its model.
Built with what exists
NCS model weights and code; the Netherlands F3 facies benchmark, an open labelled 3D geological model with baseline code and data on Zenodo; one graphics card; Python.
First step
Download the F3 benchmark and the NCS weights, and reproduce the F3 baseline score.
Done looks like
Scores for the NCS models on the F3 benchmark's own test split, set beside the benchmark's published baselines, with code that a second group re-runs to within one percentage point.