AI for efficient geotechnical soil investigations

August 22nd, 2019 22. August 2019

During the 17thEuropean Conference on Soil Mechanics and Geotechnical Engineering in Reykjavik, Iceland, we give insight into the inner workings of integrated soil investigations programs based on integrating geotechnical and geophysical data with machine learning.

Assessing geological risk is a significant part of planning; however, this risk is hard to control given the high cost of detailed ground investigation programs using traditional approaches, namely geotechnical drillings. 

Infrastructure cost overruns and delays are persistent challenges for engineers and project owners. Reported average cost overruns from 20 to 50 % for linear infrastructure are typical.

Helicopter Geoscanning in Mid-Norway supporting Nye Veier AS highway planning


Valuable models

Airborne geoscanning is a technology that is increasingly being used to mitigate geological uncertainty. EMerald Geomodelling has successfully translated complicated geophysical models to parameters valuable for engineers using artificial neural networks.

In the article, Large scale, efficient geotechnical soil investigations applying machine learning on airborne geophysical models, the authors illustrate the applicability of airborne geoscanning surveys to derive bedrock topography or depth of cover. 

“We had highly variable bedrock topography and challenging soil conditions, and many difficult design choices, including both alignment and whether to tunnel or above ground. Having a geoscanning-based bedrock model early in the project phased helped us in prioritizing ground investigations and help us narrow down road design options much more efficiently, says Sigbjørn Rønning, Senior Engineer at the Norwegian Geotechnical Institute.

Cost reductions

The key to good results is tight integration of accurate geophysical models and sparse geotechnical data. Doing so leads to bedrock topography models with an uncertainty of a few meters or 20 – 30 % of the sediment thickness.

Bedrock elevation models based on geoscanning coverage combined with 3, 25 and more than 1000 boreholes 

EMerald Geomodelling has tested the performance of neural networks both on computer-simulated test data as well as road and railway projects with extensive validation data. Studying the accuracy of bedrock models based on a handful to over one thousand in-situ drillings revealed the achievable accuracy and possible cost savings in infrastructure planning phases. 

“With the help of the integrated bedrock model right at the start of our project, we omitted about 200 of the originally planned 600 boreholes,” says Rønning, referring to a 2019 road design project in Norway.

The desired accuracy, which depends on the project phase, governs the savings potential. Drilling costs reductions of 20% to 50% are typical and come in addition to time savings.