Integrating multiple geotechnical data types with machine learning to construct high-resolution 3D geological models

In the central and western parts of the Netherlands, the low-lying areas are vulnerable to flooding by rivers. During times of peak runoff, dikes are essential to keep the land dry and the people safe. Rigorous safety standards are in place to ensure dikes are capable of withstanding extreme water level conditions. Key components for the strength and stability of a dike are the internal structure and composition of the dike and the geology in the subsurface: a sandy aquifer may lead to piping and undercutting of the dike while weak or layered strata under certain hydraulic pressures could potentially lead to collapse and catastrophic failure of the dike.

For the dike reinforcement project ‘Sterke Lekdijk’, the regional water authority ‘Hoogheemraadschap de Stichtse Rijnlanden’ is investigating a 55 km long section of the dike along the right bank of the river Lek. Detailed knowledge about the subsurface is crucial when quantifying the conditions of dikes. Given the very heterogeneous build-up of the Holocene sediments this is not an easy task. For the shallow subsurface (down to 50 m below surface level) TNO – Geological Survey of the Netherlands builds and maintains a nation-wide stochastic 3D geological model called GeoTOP. With a 100x100x0.5 m voxel size this model gives a sense of the overall geology, but lacks the very detailed information below the dikes that is needed for the task at hand.

Construction of a high-resolution geological model requires a high data density. Traditionally, shallow geological models are based on borehole information. However, in the built environment another data source is available in the form of cone penetration tests (CPTs), which are routinely obtained to measure the strength of subsurface sediments for geotechnical purposes. Although classification charts are available to translate CPT measurements into lithological classes, these charts require adjustments for local use and resulting performance remains variable. To enable the use of CPTs for geological modelling an artificial neural network (ANN) was trained to translate CPT measurements to lithological classes. Training of the ANN was done on neighboring borehole-CPT pairs (spaced at max. 10 meters). The ANN produces realistic results, with cross-validation statistics showing a vast increase in performance of the ANN results compared to traditional classification charts.

The disclosure of CPTs for geological modelling greatly increases the data density along man-made structures such as dikes. A local high-resolution version of the GeoTOP model was constructed, with a voxel size of 25x25x0.25 m. This detailed information includes the lithostratigraphical unit the voxel belongs to, the most probable lithological class of the voxel as well as the probability of occurrence of particular lithological classes. The high-resolution model enables the local water authority to better estimate dike stability, better target additional measurements in areas of high uncertainty, and take more location specific reinforcement measures.