Application of Physics-Informed Machine Learning to Geotechnical and Geophysical Site Investigation Data To Define Centimetre-Scale Design Parameters for Offshore Wind

Abstract

Offshore wind plays a pivotal role in enhancing Europe's energy security and achieving energy decarbonization goals. However, expediting offshore wind deployment necessitates efficient and economical site investigation surveys. To address this challenge, we introduce a novel approach utilising a deep neural network (DNN) to establish correlations between geotechnical cone penetrometer test (CPT) data and shear wave velocity (𝑉) from seismic CPT. Subsequently, porosity and P-wave velocity (𝑉) are derived using a 𝑉 to bulk density correlation and a dynamic poroelastic model. The DNN is trained and tested on a dataset comprising 5284 instances of public-domain geotechnical CPT test data, including depth, tip resistance, sleeve friction, and 𝑉 from seismic CPT. During testing, the DNN model demonstrates a mean absolute error of 55 m s-1 between predicted and measured 𝑉 values. The uncertainty in 𝑉 predictions is attributed to factors such as (i) limited training data for some soil types such as gravelly sands, (ii) intricate relationship between geotechnical CPT features and seismic properties influencing 𝑉, (iii) the presence of CPT features and 𝑉 combinations that lie well outside the region from most combinations (i.e. outliers), and (iv) CPT features and 𝑉 measurements that are averaged over different depth ranges. The derived porosity and 𝑉

Full Paper