Accelerating Geotechnical Soil Characterization in Offshore Windfarm Sites via Semi-Supervised Learning

Summary As an essential component to the process of wind farm site characterization and selection, robust soil property estimation becomes feasible by integrating the available 2D ultrahigh-resolution (UHR) seismic profiles and 1D come penetration tests (CPT) using machine learning algorithms. However, the strong noises present in UHR seismic and moreover limited availability of CPT logs increases the risk of unstable property estimation by typical supervised learning. This study proposes developing a semi-supervised learning workflow for geotechnical soil characterization, which consists of two steps: seismic denoising and feature engineering (SDFE) and seismic-CPT integration (SCI). Each of them is implemented by training a deep convolutional neural network (CNN) and they are connected by using the encoder of the pre-trained SDFE-CNN as the basis of the SCI-CNN. The proposed method is tested on the Hollandse Kust Zuid (HKZ) wind farm between the Hague and Zandvoort. The machine prediction successfully delineates the sandy silt layer of low friction ratio and medium cone-tip resistance below the seabed and moreover the underlying potential clay intervals of relatively high friction ratio and low cone-tip resistance.