top of page

AI in geotechnical work

Nov 5
3 min read

Updated: Nov 6

AI in geotechnical work usually means one of two things. One is a general tool like ChatGPT used on everyday documents. The other is a machine learning model trained on ground data. They have different uses and different risks, so it's worth keeping them apart.


Engineer in a hard hat and hi-vis vest looking up at a roadside slope with marker pegs and a solar-powered sensor
AI-generated image.

What Engineering New Zealand's guidance says

Engineering New Zealand's guidance is clear on where the general tools stop. It describes AI as an assistant and not a decision-maker. Its geotechnical example is that AI can help draft summaries of site investigation logs, but can't interpret field or lab test results or replace geotechnical modelling. More broadly, it says AI can't sign off work or replace site investigation, inspection or peer review. Engineers remain liable under the Building Act, the Health and Safety at Work Act and the Engineering New Zealand Code of Ethical Conduct.


Where general AI tools can help

That still leaves useful work. Drafting a summary from a set of logs is a reasonable place to start, as long as someone who knows the job checks it against the originals. Take care with standards. The guidance warns that AI may invent clauses or misquote data, and it says not to upload NZS documents to AI tools, because copyright stays with Standards New Zealand and uploading needs a specific licence. Building Code extracts can be uploaded, with the clause reference and revision date. The guidance states it was current as of July 2025, so check for changes.


Site reports need thought too. Check your engagement terms before putting client reports into an outside tool. We've written more about that in our article on keeping client data private.


Machine learning on ground data

Machine learning on ground data is a different kind of work, and New Zealand has a lot of data for it. The Natural Hazards Commission says the New Zealand Geotechnical Database holds test data worth an estimated more than $500 million, and has more than 7,000 users. Full records need registration and acceptance of its terms.


Liquefaction is where we found the clearest New Zealand examples. A 2025 NZSEE conference paper by Jacka, Maxim and Russel, funded by Toka Tū Ake EQC, trained a neural network on full CPT profiles, using over 47,000 case histories from ten New Zealand earthquakes. It aims to correct the Liquefaction Severity Number where it overpredicts in complex interbedded soils. Sanger, Geyin and Maurer, in a paper posted to arXiv, built a New Zealand model from 16,475 CPT locations. They say their models are best suited to regional-scale hazard simulation and near-real-time response. That's useful for planning and for the hours after an earthquake, but it doesn't replace testing on your site.


Landslide mapping after Cyclone Gabrielle

Landslide mapping shows why people still check this work. After Cyclone Gabrielle, GNS Science and its partners mapped more than 140,000 landslides, using nearly 17,000 aerial photos across 50,000 square kilometres. NZ Herald reported in September 2023 that automated mapping missed as much as 30% of the landslides when it was tested, and that GNS's Chris Massey said this supported having people map the event.


Measuring ground movement

Our own work is on the measurement side. Our LandSure reports set out what satellite radar measured at a property, and an engineer is the person to explain why the ground is moving. We've also written about monitoring land stability with sensors and satellites.


How to test a tool

If you want to try a tool, test it on a finished job where you already know the answer and see what it gets wrong. In an article for Engineering New Zealand, Daniel van der Walt and David Dempsey of the University of Canterbury say engineers need to judge whether AI is competent for a given task, verify what it produces and keep a person in the loop. For geotechnical work, that person needs to be someone who can spot a wrong answer.


Sources

Sources: Engineering New Zealand, "AI in professional practice", stated current as of July 2025. Daniel van der Walt and David Dempsey, "AI and engineering competence", Engineering New Zealand, 1 May 2025. Natural Hazards Commission, New Zealand Geotechnical Database. Jacka, Maxim and Russel, "National Liquefaction Model: framework, opportunities and challenges", NZSEE 2025 Annual Conference, listed by Tonkin + Taylor. Sanger, Geyin and Maurer, "Mechanics-Informed Machine Learning for Geospatial Modeling of Soil Liquefaction", arXiv 2509.10962. Jamie Morton, "140,000 landslides: What scientists are learning from Gabrielle's record damage", NZ Herald, 23 September 2023.


Comments


bottom of page