Industries / Construction and property / Heavy civil / Tunneling
Ground investigation
Ground investigation finds out what rock, soil, water and gas lie along a tunnel's route, before and during digging, and watches how the ground above moves while the tunnel is built.
5
Milestones achieved, from 1908 to 2019.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1908
Lötschberg breaks into gravel
The Lötschberg rail tunnel in Switzerland was planned to pass under the Gastern valley through solid rock. On 24 July 1908, about 2,675 m from the north portal and some 180 m below the valley floor, a blast at the face opened into loose debris that poured in and filled the tunnel for about 1,800 m, by the account the geologist Albert Heim published that year. Heim counts 24 workers buried; the Swiss broadcaster SRF counts 25 miners and four horses. The flooded stretch was walled off for good, and the tunnel was rerouted on an S-shaped curve that SRF says added about 800 m. SRF reports that a Zurich geologist had warned that the exploratory borings were not deep enough.
Lessonthe depth of sound rock under a valley has to be proven by drilling, not assumed.
Credit
the Berner Alpenbahn-Gesellschaft (builder); Albert Heim (account)
1964
Counting sound core
Rock quality designation (RQD) is the share of a run of drilled rock core that comes out in sound pieces longer than 100 mm (4 in). Don Deere devised it in 1964, while working on granite sites at the Nevada Test Site, as a quick way to grade rock from the cores that site investigations already drill. It became part of the two main rock mass ratings, RMR and the Q-system. In 2017 Pells, Bieniawski, Hencher and Pells reported that 17 independent professionals mapping rock in Australia and South Africa gave RQD values different enough to cause significant errors in RMR and Q, and that many parts of the world no longer apply Deere's full definition.
Lessona simple index drifts when people stop applying its exact definition.
Credit
Don U. Deere
1974
Six numbers rate rock
The Q-system grades the rock around a tunnel with six numbers taken from mapping or core: RQD, the number of joint sets, joint roughness, joint alteration (clay or weathering in the cracks), water, and a stress factor. Combined into one Q value, it is read against a chart of the support used before in similar rock, such as rock bolts and sprayed concrete. The Norwegian Geotechnical Institute (NGI) developed it from 1971 and published it in 1974. NGI revised the support chart in 1993 from 1,050 recorded cases, mostly Norwegian, and in 2002 from more than 900 more from Norway, Switzerland and India, and issued a revised handbook in 2025. Pells and colleagues describe RMR and Q as playing "substantial roles" in rock design worldwide.
Credit
N. Barton, Lien and Lunde, NGI
1997
Grout holds Big Ben
London's Jubilee Line Extension dug a station box up to 39 m deep and two 7.4 m platform tunnels beside the Big Ben Clock Tower at Westminster. The engineers' analysis showed the tower would otherwise tilt by at least 120 mm at the height of the clock faces. The answer was compensation grouting: grout, a mix of cement, sand and water, is pumped through a fan of steel tubes about 50 m long laid under the foundation, in step with the digging, to make up for ground that moves toward the excavation. Tilt reached the 15 mm trigger, measured 55 m up, in December 1995. From January 1996 to September 1997 crews injected 122 m³ of grout in 24 episodes and held the tilt in the 15 to 25 mm range that had been set.
Credit
Harris, Mair, Burland and Standing (advisers); Balfour Beatty and AMEC (contractor)
2019
Settlement measured from orbit
Satellite radar interferometry (InSAR) compares radar images of the same ground taken at different times to measure movements of a few millimeters at thousands of points. Giardina and colleagues compared InSAR readings over London during the digging of Crossrail with conventional survey readings, then used the radar data to work out how individual buildings deformed. Conventional monitoring covers few buildings because of its cost; the radar covered whole districts. They published the method, with its data, in January 2019.
Credit
Giardina, Milillo, DeJong, Perissin and Milillo
Watch for
Strata predicted from machine data
A shield tunneling machine logs its thrust, torque and speed as it digs. Tao Yan, Shui-Long Shen, Zhen-Yu Yin and Ning Zhang combined these logs with boreholes, the holes drilled along the route before digging, to redraw the ground layers between the boreholes. Checked against real boreholes held back from the model, their method raised the accuracy of the predicted cross-section from 0.77 to 0.938, and they applied it on a shield tunnel in Guangzhou, China. A second team, Xiangcou Zheng and colleagues, reports more than 95% accuracy on synthetic 3D cases with a method that updates a borehole model as machine data arrive, and applied it on the Guanghua intercity tunnel.
Achieved whena team not involved publishes a forecast of the ground ahead before a machine reaches it, on another project, and the ground actually met is then mapped and compared.
Claimed by
Yan, Shen, Yin and Zhang, in Automation in Construction, January 2026; Zheng and colleagues, 2026
Open
01
Face ratings from many geologists
A public set of 200 photos of tunnel faces and rock exposures, each rated blind on the Q-system's six numbers by at least five engineering geologists, and a test of whether AI vision models land inside the spread between the people.
The briefCredit · unclaimed02
Settlement forecasts checked from orbit
A public table that sets, for at least 10 tunnel boring machine drives in at least four European countries, the ground loss the designers assumed beside the ground loss measured from free satellite radar. Ground loss, or volume loss, is the volume of extra ground that moves into a tunnel as it is dug, given as a percentage of the tunnel's own volume; it sets how deep the dip in the surface above will be.
The briefCredit · unclaimed
Open 01 · the brief
Face ratings from many geologists
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of 200 photos of tunnel faces and rock exposures, each rated blind on the Q-system's six numbers by at least five engineering geologists, and a test of whether AI vision models land inside the spread between the people.
Why it is still open
Obayashi's deep-learning system matched its engineers' judgments of mountain tunnel faces 73% to 97% of the time on the company's own records, and those records are not public. The open-source TRaiC toolkit maps faces from 360° photos and scores RMR, but ships no set of photos rated by several people. Pells and colleagues showed in 2017 that 17 professionals disagree enough on RQD alone to shift RMR and Q, measured on surface rock. In a 2003 study at the Norwegian University of Science and Technology (NTNU), Bjørn Nilsen and four colleagues each mapped the same three rock exposures along a planned tunnel near Trondheim; their individual ratings differed significantly, though the final rock classes and support came out quite similar. No public set of face photos with several independent ratings was found.
Built with what exists
Face photos used with permission from tunnel owners and university groups; NGI's free Q-system handbook and its rating tables; a rating form; Python for Krippendorff's alpha, a score of agreement among several raters that discounts chance; a vision model that returns probabilities for fixed answers (see the Note "A thousand questions, one pass"); TRaiC for tracing joints.
First step
Get permission to use 30 face photos from one tunnel owner or university, and have three geologists rate them blind.
Done looks like
200 photos with at least five independent ratings each, published under an open license with an agreement score for each of the six numbers; for at least two AI models, the share of photos where the model's Q value falls within the range of the human ratings.
Open 02 · the brief
Settlement forecasts checked from orbit
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table that sets, for at least 10 tunnel boring machine drives in at least four European countries, the ground loss the designers assumed beside the ground loss measured from free satellite radar. Ground loss, or volume loss, is the volume of extra ground that moves into a tunnel as it is dug, given as a percentage of the tunnel's own volume; it sets how deep the dip in the surface above will be.
Why it is still open
Studies so far take one project at a time. Giardina and colleagues compared radar with survey readings for Crossrail in London; Reinders and colleagues at Delft University of Technology checked radar against levelling over Amsterdam's North/South metro line (2021); Rigamonti and colleagues used the European Ground Motion Service (EGMS) to back-calculate a mean volume loss of 0.33% for twin machine-bored tunnels in river deposits (2025); Rivas and colleagues fitted the surface dip over a Taipei metro tunnel from Sentinel-1 radar (2026). No study was found that lines many tunnels up against their designers' own assumptions.
boa.unimib.itisprs-annals.copernicus.orgresearchportal.bath.ac.ukrepository.tudelft.nl
Built with what exists
EGMS, the Copernicus service that publishes Sentinel-1 ground movement for most of Europe free of charge; the volume loss assumed in each project's published environmental statement or design papers; the Gaussian settlement trough, the bell-shaped dip engineers fit to surface settlement; Python with GeoPandas; a web agent to gather design documents (see the Note "Deleting 80% to start over").
First step
Download EGMS data over one machine-bored drive finished during the years EGMS covers, and fit a settlement trough to one cross-section.
Done looks like
At least 10 drives, each with the assumed and the radar-measured volume loss and trough width with their uncertainties, published with data and code, and the share of drives where the measured loss came in below the designers' assumption.