Introduction
The goal of this project is to develop a mechanism-based machine learning algorithm, trained using a large database of field data, to optimise the installation of offshore wind turbine foundations and shape the decision-making process during installation.
In the UK alone, thousands of anchors and foundations must be installed every year to support the (floating) offshore wind turbines necessary to achieve the net zero objectives. Piles are currently the most commonly used foundation type and pile driving is the main installation method. However, this pile hammering creates loud underwater noise which is particularly detrimental to marine mammals. On the contrary, suction piles are installed by pumping water on the inside of the pile, which creates a suction effect that pushes the pile into the ground without any impact noise.
The installation of suction piles is challenging. Premature refusal, which is the termination of the installation before reaching the targeted depth, may require the full removal of the pile before its relocation, which has a significant (carbon) cost. Refusal can be due to mechanisms such as piping (breaking the internal suction) or plug heave (internal soil upwards movement), but they are not yet well predicted, especially in challenging geologies such as hard or layered soils. Innovative techniques, such as suction cycling, have been recently introduced to overcome those issues, but their effect is not yet predictable.
The goal of this project is to develop a mechanism-based machine learning (ML) algorithm, with associated uncertainty quantification (UQ), to optimise the installation of suction pilesin terms of site selection or installation parameters (pumping sequence) to reduce installation time and risk. Standard ML may produce accurate predictions but will not typically account for engineering and physical knowledge of the suction process. Greater efficiency, in terms of required data and improved UQ, may be obtained by constraining the ML model via the known physical mechanisms. Such an ML model may also allow for limited extrapolation away from the training data. An effective model of the process can then be interrogated to propose new installation strategies and locations.
This goal will be achieved by meeting the following research objectives:
Identify the different failure mechanisms encountered during suction pile installation and determine their geotechnical features.
Develop a mechanism-based predictive model for suction pile installation that accounts for uncertainties in soil properties and provides model predictions that enable live decision making in the field. A first attempt might be using a physically constrained Gaussian process.
Propose new installation strategies, e.g, using techniques based on Bayesian optimisation, and quantify their effect on suction pile installation.