Comparison of Machine Learning Models in a Data-Driven Approach for Scalable and Adaptive Design of Laterally-Loaded Monopile Foundations-Paper

Paper

Pdf Summary

The paper proposes a data-driven approach to designing laterally loaded monopile foundations, aiming to combine the efficiency of rule-based PISA models with the site-specific accuracy of models calibrated through bespoke 3D finite-element analyses. Instead of relying on predefined soil-reaction rules, the approach trains machine-learning models on finite-element data to predict soil-reaction parameters for different pile geometries and soil profiles. As the database grows, the model can incorporate past design experience and reduce the need for new, resource-intensive analyses.<br /><br />The study compares polynomial regression, neural-network (NN) regression, and Gaussian-process (GP) regression. The models are trained and tested using data from analyses of two monopile geometry ratios in Cowden till, with inputs describing depth and pile geometry and outputs representing soil-reaction parameters.<br /><br />Polynomial regression is the easiest to interpret but performs least accurately and struggles to capture strongly nonlinear patterns. NN regression is more accurate but harder to interpret. GP regression provides the best overall accuracy and remains reasonably interpretable: its predictions draw on nearby data points, while its confidence intervals indicate where predictions are uncertain.<br /><br />In a design example, all three data-driven models closely reproduce the ground-level load–displacement behaviour predicted by the numerical PISA model; GP gives the closest agreement. The authors therefore identify GP regression as the most promising option. Its uncertainty estimates could help identify when additional finite-element analyses are needed to reduce design risk, while its capacity to adapt as data accumulate offers a scalable route to future designs.<br /><br />The findings are based on a limited dataset for one soil profile and two geometry ratios. Further data and validation across a wider range of soils and conditions would be needed to establish broader applicability.

Keywords

monopile foundations

data-driven design

machine learning

finite-element analysis

PISA model

soil-reaction parameters

Gaussian-process regression

neural-network regression

polynomial regression

lateral load-displacement behavior