By Jamie Alexander
In an era of AI generated music, images, and poems, could we be on a path towards the first AI generated full wind turbine design? Or more pragmatically, can machine learning (ML) be used to automate and optimise geodata interpretation and foundation design? I was able to investigate these ideas in-depth as part of my Master’s thesis, during a seven month placement with Geowynd.
The project explored two areas for implementation of ML to the geotechnical interpretation and design process. For Scope A), I developed an ML tool capable of predicting shear wave velocity (Vs) from CPTs, aka the ‘Virtual SCPT’*; Scope B) involved training a ML model to predict the response of monopile supported offshore wind turbines. The results of B) were remarkably impressive, providing finite element analysis (FEA) quality output at a fraction of the time.
The model developed in B) was trained on thousands of randomised soil profiles and proved to be capable of predicting a wind turbine monopile response with promising accuracy when compared to high fidelity FEA as shown below:
The methodology of B) is confidential (for now!) and hence I have focused on giving a brief overview of A) below.
To develop the Virtual SCPT (i.e. Scope A), an ML model was trained using XGBoost, which implements an algorithm based on the basic concept of decision trees. By making accurate predictions of Vs, dependence on a high number of slow and costly (real) SCPT tests can be reduced at an offshore windfarm. Open-sourced geo-data was available to train and test the model thanks to the Netherlands Enterprise Agency (RVO). The model was then validated on two locations from the same windfarms as the training data; however, these locations were excluded from the dataset prior to training. Therefore, they acted as a ‘blind’ prediction. It was found that the model was able to successfully predict the variability of Vs throughout the entire depth of both locations.
XGBoost has an advantage over conventionally used empirical Vs-CPT correlations in that it can be continuously updated as new data becomes available. Therefore, there is certainly scope for a machine learning model to replace (or enhance) current empirical methods of rapidly estimating Vs directly from CPT data.