Published on

The Ground Is the Variable

Authors

Listen to this article · The Ground Is the Variable

0:00 / 0:00

Watch this article · The Ground Is the Variable

Metro tunnels, underground stations and deep excavations are built next to things that cannot move: buildings, roads, water mains, cables. The tolerances are millimetres. And unlike steel or concrete, the ground arrives without a specification — heterogeneous, different ten metres apart, and changed again by the digging itself.

That is why the useful AI in tunnelling is not the kind that writes reports. It is the kind that turns sparse measurements into a picture of a material nobody can see.

13,926
borehole records behind one city model
163
local soil samples in one prediction study
2
days of the Hanoi geotechnical conference

First the ground, then the machine

The most mature application is the ground model. A research group in Japan trained a neural network on 13,926 borehole records to build a three-dimensional map of soil properties beneath Tokyo. Published in 2025, the model predicts density and state at locations where no borehole exists — which is most of the city. A consulting engineer in the United Kingdom built a tool that folds monitoring data back into the geotechnical parameters themselves, and used it to choose construction methods for deep excavations in central London.

  1. Boreholes

    sparse, expensive, point measurements

  2. Interpolate

    a model fills the empty space

  3. Predict

    density, stiffness, expected settlement

  4. Decide

    foundation type, route, support method

The model is not the ground. It is the best available guess between two holes.

The second application is the machine. A tunnel-boring machine already reports thrust, cutterhead rotation and advance rate continuously, and every metre it drives is labelled with the geology it met. That is a training set nobody had to collect on purpose. A university team in China built an intelligent control layer on exactly those signals; in Vietnam, engineers point at the same three as the obvious first model — predict the advance rate, then adjust the parameters before the ground forces the adjustment.

Borehole records, one city13,926

the dense case

Soil samples, one study163

the sparse case

The same method at two scales. Scale is the difference between a model and a sketch.

The layer that is not about soil

The third application is the site itself. A housing authority in Singapore runs AI video to flag unsafe behaviour and alert supervisors, and since January 2023 contractors bidding for new public housing must apply AI video analytics. That is the most underrated mechanism in the whole trend: procurement, not research. A tender requirement moves an entire supply chain in a way a conference paper never does.

A tunnel boring machine, drawn for ABN Asia

Where the method breaks

A prediction is not a measurement. Interpolating between two boreholes produces a number with no error bar unless someone puts one there, and underground, a wrong number does not announce itself. It appears months later as settlement in a building that was fine last quarter.

The discipline that separates a model from a guess is unglamorous and identical everywhere: validate against instrumentation, not against the training set. Vietnamese researchers started early on this — a geotechnical-and-AI group formed in 2018, a 2019 model trained on 163 local soil samples to predict the consolidation coefficient of soft soil, which is the number that decides how much a road will sink over the next decade. The international society has since formed a technical committee on machine learning and big data, and a Hanoi conference this month put the same question on the table: what does a prediction mean when the ground is the only thing that gets a vote.

What actually changes

The trend is not that AI is entering construction. It is that the three things tunnelling always lacked — dense ground data, a machine that logs itself, and a procurement rule that forces adoption — arrived in the same decade. Where all three exist, the model is useful. Where one is missing, the model is a slide.

Underground, being wrong is measured in millimetres, and the software that matters is the part that notices. You not only get software — you get secure software. If you are putting predictions into a decision that can move a building, talk to us: [email protected].

Source: VnExpress, 20 November 2026.