Industries / Construction and property / Heavy civil
Tunnels
Tunneling digs passages through ground or rock, or under water, and lines them so that they stay open and dry.
7
Milestones achieved, from 1843 to 2016.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
In depth
The same ground is charted in more depth, as an industry of its own.
The road
AchievedWatch forOpen
Achieved
1843
A shield under the Thames
Marc Brunel's tunneling shield was an iron frame, pushed forward by screw jacks, that held up the ground around and in front of the miners while they dug. Work on a tunnel under the River Thames in London began in 1825, and it opened on March 25, 1843, after floods, gas and sickness stretched the job to 18 years. The newsletter Construction Physics calls it "the first tunnel under a body of water in the world"; Wikipedia says "the first tunnel known" under a navigable river and notes a possibly older Babylonian tunnel under the Euphrates.
Lessonthe shield made soft ground under a river workable, but floods still broke in and the work advanced only about 8 ft a week.
Credit
Marc Isambard Brunel; Isambard Kingdom Brunel; Henry Maudslay (built the shield)
1871
Air-powered drills cross the Alps
The Fréjus (Mont Cenis) rail tunnel between France and Italy is 13.7 km long. Its crews drilled the holes for explosive charges with machines driven by compressed air, ten drills mounted on one frame. The two crews met on December 26, 1870, and trains ran from September 17, 1871.
Lessonmachine drilling shortened the job to about 14 years; at the old hand-drilling rate it would have taken "at least forty" years by an 1871 newspaper's estimate, or 71 years by Wikipedia's.
Credit
the Italian and French project teams under engineer Germain Sommeiller
1952
A machine bores through rock
A tunnel boring machine (TBM) is a full-width rotating cutting wheel with the machinery to push it forward and carry away the spoil. The Robbins Company says that "in 1952 James Robbins developed the first modern Tunnel Boring Machine for the Oahe Dam Project in South Dakota." It dates its first rock tunnel cut with disc cutters alone to 1956, on the Humber River in Canada. Construction Physics dates the Oahe machine to 1954 and says it advanced almost 200 m a week, about five times faster than drill and blast.
Credit
James S. Robbins, The Robbins Company
1974
Soil pushes back on soil
An earth pressure balance (EPB) shield keeps the soil it has just dug in a sealed chamber behind the cutting wheel, at a pressure that holds back the ground and groundwater in front. This lets machines dig through soft, wet ground without it flowing into the tunnel. A Japan International Cooperation Agency report says "the EPBS TBM was invented in 1974 and has been developed in Japan," with more than 1,400 projects using it in Japan and other countries.
Credit
the Japanese tunneling industry; no source found names a company
1994
Heathrow tunnels collapse
The New Austrian Tunneling Method digs the ground in stages, sprays a thin concrete lining onto it at once, and measures how the ground moves to decide what support to add. On October 21, 1994, station tunnels being dug this way for the Heathrow Express rail link in London collapsed and kept failing for days. Nobody was hurt. The UK Health and Safety Executive found that "warnings of the approaching collapse were present from an early stage in construction but these were not recognised." Contractor Balfour Beatty and designer Geoconsult were fined a total of £1.7 million.
Lessonmeasuring ground movement protects nobody unless someone reads the readings and acts on them.
Credit
UK Health and Safety Executive (investigation)
2013
One machine stops a tunnel
Bertha, a 17.5 m TBM built by Hitachi Zosen, was boring Seattle's SR 99 highway tunnel when it overheated and stopped in December 2013. Crews found broken bearing seals and dug a 120 ft deep, 80 ft wide shaft in front of the cutting head to lift it out for repair. About two years of repairs followed, and the machine broke through at the end of its 1.75-mile drive in April 2017. Who was to blame went to court; the contractor sued a soil-testing firm over an 8-inch steel pipe left in the ground.
Lessonwhen a whole tunnel rests on one machine, any fault in it stops all the work.
Credit
Washington State Department of Transportation (owner); Seattle Tunnel Partners (contractor); Hitachi Zosen (machine)
2016
A 57 km base tunnel
The Gotthard Base Tunnel runs 57.1 km under the Swiss Alps. The first blast was on November 4, 1999, and the tunnel opened on June 1, 2016. Wikipedia says four Herrenknecht TBMs bored about 45 km of each of its two tubes, and calls it the world's longest railway tunnel.
Credit
AlpTransit Gotthard AG (builder); Herrenknecht (machines)
Watch for
Fehmarnbelt immersed tunnel
An immersed tunnel is built on land in giant concrete sections, which are floated out, sunk into a trench dug in the seabed and joined. Femern A/S says its 18 km road and rail tunnel between Denmark and Germany "will be the longest in the world." By September 2026 four of its 89 sections had been placed; each standard section is 217 m long and weighs more than 73,500 metric tons. On November 13, 2026, it postponed the next placement because the rubber seal on the last section placed was damaged, and said it did not yet know how long the repair would take.
Achieved whenall 89 sections are placed and sealed and the tunnel carries traffic.
Claimed by
Femern A/S, September and November 2026
Robots that change cutters
Disc cutters on a TBM wear out and are swapped by hand: workers climb into the cutting head and move cutters that can weigh 191 kg each. China Railway Engineering Corporation says the two TBMs it built for Sydney's Western Harbour Tunnel carry a robot that changes cutters with nobody in the loop, 5 to 8 times faster than manual changes. The report gives no field data.
Achieved whenthe tunnel's owner, Transport for NSW, or another outside party publishes how many cutter changes the robots made and the hours each took, compared with crews.
Claimed by
China Railway Engineering Corporation, August 2026, as reported by a general-news site
Open
01
Many analysts, one TBM log
A test of how much results depend on the analyst. A TBM logs thrust, torque and speed for every push forward, and engineers sort each push into regular or exceptional advance. Give the three open TBM datasets released with a 2025 paper by Erharter and colleagues to at least five independent analysts and two AI coding agents. Each builds its own processing steps from the paper's text alone, without the reference code. Publish every result and how often each agrees with the authors' benchmark.
The briefCredit · unclaimed02
Warnings before tunnel collapses
From the open TU Dortmund database of tunnel collapses, take 50 collapses during construction that have a public investigation report or a detailed published account. For each, record whether warning signs such as measured movement, cracking or water inflow were reported before the collapse (yes, no or not stated), how long before, and whether anything was done, with the exact sentence and source. Give the same 50 sources to two AI models and score their answers against the hand-checked key. Add at least 10 collapses from 2020 to 2025 that the database does not yet cover.
The briefCredit · unclaimed
Open 01 · the brief
Many analysts, one TBM log
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of how much results depend on the analyst. A TBM logs thrust, torque and speed for every push forward, and engineers sort each push into regular or exceptional advance. Give the three open TBM datasets released with a 2025 paper by Erharter and colleagues to at least five independent analysts and two AI coding agents. Each builds its own processing steps from the paper's text alone, without the reference code. Publish every result and how often each agrees with the authors' benchmark.
Why it is still open
The paper says the result depends on the analyst's choices and publishes only one set of processing steps: "Even slight deviations from one commonly agreed upon processing pipeline can lead to different advance classifications." Its data are synthetic because "in geotechnical engineering, confidentiality constrains possibilities for the use of real datasets." No comparison of many analysts was found.
Built with what exists
The MIT-licensed repository github.com/geograz/TBM_advance_classification, with three synthetic TBM datasets and a benchmark script; Python with pandas and scikit-learn for Cohen's kappa, an agreement score that discounts agreement by chance; two AI coding agents. Coding agents that can turn a methods section into working code are what make a many-analyst test cheap.
First step
Clone the repository, run the benchmark on the first dataset, check the counts against the README, and write the one-page brief each analyst will get.
Done looks like
A public table of at least 7 analyses by 3 datasets, each with the share of pushes classed the same as the benchmark and a kappa score, plus all code and a one-page note on which processing step caused most of the disagreement.
Open 02 · the brief
Warnings before tunnel collapses
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
From the open TU Dortmund database of tunnel collapses, take 50 collapses during construction that have a public investigation report or a detailed published account. For each, record whether warning signs such as measured movement, cracking or water inflow were reported before the collapse (yes, no or not stated), how long before, and whether anything was done, with the exact sentence and source. Give the same 50 sources to two AI models and score their answers against the hand-checked key. Add at least 10 collapses from 2020 to 2025 that the database does not yet cover.
Why it is still open
The Heathrow inquiry found that warnings were present early but not recognized, yet the open database has no field for warning signs. Of its 415 records, from 1941 to 2019, only 201 have any text in the column for incidents, reports and causes.
Built with what exists
The database spreadsheet (DOI 10.17877/DE290R-23148, open access); public investigation reports and news archives; two AI models that read long PDFs, with a reviewer model that flags doubtful answers for a person (see the Note "Seven more flagged"); a spreadsheet.
First step
Download the spreadsheet, filter to construction-stage collapses that name a source, and code the first 10 by hand to fix the rules before any AI model sees them.
Done looks like
A public table of 50 coded collapses, each with a source quote checked by a person; percent agreement with the key for each AI model; 10 or more new cases from 2020 to 2025, with sources.