Machine Learning and Quantitative Ground Models for Improving Offshore Wind Site Characterization

Quantitative integrated ground models are a requirement for proper cost optimal site characterization, for offshore renewables, coastal activities and O&G projects. Geotechnical analyses and interpretations often rely on isolated 1D boreholes. On the other hand, geophysical data are collected in 2D lines and/or 3D volumes. Geophysical data therefore provides the natural link to re-populate geotechnical properties found in the 1D boreholes onto a larger area and thereby build a consistent and robust ground model. The geophysical data can be used to estimate geotechnical data and, as of today, there are a few methods available that can reliably map the dynamic properties from the seismic data (stratigraphic information, P-wave velocities, amplitudes, and their attributes) into geotechnical or geomechanical properties, particularly for shallow sub-surface depth. Being able to predict soil properties away from boreholes is important, as often the field layout changes during the development phase, and hence, information at the specific foundation locations may not be readily available. We have developed a workflow to build quantitative ground models following three approaches: (i) a geometric model in which the seismic data interpretations guide the prediction of geotechnical properties; (ii) a geostatistical approach in which in addition to the structural constraints, we used the seismic velocities to guide the prediction; and (iii) a multi-attribute regression using an artificial neural network (ANN). We apply it to a set of publically available data from the Holland Kust Zuid wind farm site in the Dutch sector of the North Sea. The result of the workflow yields maps or sub-volumes of geotechnical or geomechanical properties across the development site that can be used in further planning or engineering design. In this study, we use the tip resistance from a CPT as an example. The tip resistance derived using all methods generally give good results. Validation against randomly selected CPT shows good correlation between predicted and measured tip resistance. The ANN performs better than the geostatistical approach. However, these two approaches require good data quality and a rather large dataset to be effective. Therefore, using a global dataset not restricted to the Holland Kust Zuid site may improve the prediction. Moreover, using existing empirical correlation and calibration through laboratory testing or by training another ANN model, the geotechnical stiffness/strength parameters such as angle of friction or undrained shear strength could be derived. The next step is to use the results and their uncertainty into a cost assessment for the given foundation concepts.