Estimating geotechnical parameters using semisupervised learning: An example from the Dutch offshore wind farm zone

Abstract

Wind energy is considered to be of great importance for promoting energy transition and achieving net-zero carbon emission. Reliable modeling and monitoring of the near subsurface geology is crucial for successful wind farm selection, construction, operation, and maintenance. For optimal characterization of shallow seafloor sediments, two-dimensional (2D) ultrahigh-resolution (UHR) seismic survey and one-dimensional (1D) cone-penetration testing (CPT) are often acquired, processed, interpreted, and integrated for building three-dimensional (3D) ground models of essential geotechnical parameters such as friction. Such a task faces multiple challenges, including limited CPT availability, strong noise contamination in UHR seismic, and heavy manual efforts for completing the traditional workflows particularly acoustic impedance inversion. This study proposes accelerating the integration by a semi-supervised learning workflow with three highlights. First, it enables geotechnical parameter estimation directly from UHR seismic without impedance inversion. The second comes from the use of a pre-trained feature engine to reduce the risk of overfitting while mapping massive UHR seismic with sparse CPT measurements through deep learning. More importantly, it allows incorporating other geologic/geophysical information, such as a pre-defined structural model, to further constrain the machine learning and boost its generalization capability. Its values are validated through applications to the Dutch wind farm zone for estimating four geotechnical parameters, including cone-tip resistance, sleeve friction, pore-water pressure and the derived friction ratio, in two example scenarios, (a) UHR seismic only and (b) UHR seismic and an eleven-layer structural model. Both results verify the feasibility of data-driven geotechnical parameter estimation. In addition to the two demonstrated scenarios, the proposed workflow can be further customized for embedding more constraints, e.g., pre-stack seismic and elastic/static property models, given their availability in a wind farm of interest.