Comparison of machine learning models in a data-driven approach for scalable and adaptive design of laterally-loaded monopile foundations

Suryasentana, S. K. and Burd, H. J. and Byrne, B. W. and Aghakouchak, A. and Sørensen, T.; (2020) Comparison of machine learning models in a data-driven approach for scalable and adaptive design of laterally-loaded monopile foundations. In: International Symposium on Frontiers in Offshore Geotechnics. Deep Foundations Institute (DFI), USA. ISBN 9780976322948

[thumbnail of Suryasentana-etal-FOG-2020-scalable-and-adaptive-design-of-laterally-loaded-monopile]
Preview
 Text. Filename: Suryasentana_etal_FOG_2020_scalable_and_adaptive_design_of_laterally_loaded_monopile.pdf
Accepted Author Manuscript
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo
Download (1MB)| Preview

Abstract

The design of monopile foundations under lateral loading is typically conducted using the Winkler modelling approach. Design models incorporating this approach include the 'p-y' method specified in the design codes or the more recent 'PISA' design model. These design models have predefined soil reactions that are appropriate only for the pile and soil characteristics considered in the calibration. One of the biggest uncertainties is thus defining the appropriate soil reactions for design soils with different characteristics from the calibration cases. This paper proposes a data-driven approach, in which machine learning models are used to determine the most appropriate soil reactions directly from data. As these models can automatically adapt and increase in complexity to new data, it provides for a more scalable approach than existing predefined design models. Three machine learning models (polynomial regression, neural network regression (NN) and Gaussian Progress (GP) regression) are compared. The results demonstrate that the polynomial regression model is the simplest to interpret but is the least accurate. The NN regression model has high accuracy, but is the hardest to interpret. The GP regression model is the most accurate and is reasonably easy to interpret. Moreover, the GP regression model has the additional benefit of providing intrinsic uncertainty estimates, making it appropriate for quantifying and reducing design risks.

CORE (COnnecting REpositories)