Back to search

University of Sheffield

Physics-informed machine learning for structural dynamics: combining physics and data for offshore structures

Abstract

dc:description.abstract

Physics-Informed Machine Learning (PIML) aims to exploit the benefits of both physics and data-based modelling approaches; insight, structure and an enhanced ability to extrapolate are provided through physical knowledge, whilst a data-based component increases flexibility and allows for the capture of complex relationships directly from data. For many applications within engineering, our available physical knowledge might only be sufficient to represent part of a complete systems’ behaviour, while the accompanying use of measured data can help capture variation due to environment, manufacturing tolerances, or other effects. When working effectively, PIML models often outperform the individual physics and data-based models from which they are constructed. This thesis develops PIML models within the field of structural dynamics, with a focus toward Structural Health Monitoring (SHM), offshore structures and wave loading prediction. In harsh environments, including those offshore, conditions are often difficult to fully characterise with purely physics-based approaches, whilst a high variability of conditions places large demands on the collection of measured data. Typically, the predictions of machine learners are only suitable within the realms of previously observed conditions, requiring extensive, and expensive, monitoring campaigns. PIML has the potential to address both of these issues. Amongst the thesis contributions, is the first instance of a physics-informed model for the prediction of wave loads on a real offshore structure. The widely used Morison’s Equation, an empirical wave loading solution, is incorporated within the mean function of an autoregressive form of Gaussian process Regression (GPNARX). The model achieved a 29.13% and 5.48% relative reduction in error over Morison’s Equation and a purely data-based GP-NARX respectively. Enhanced improvement was seen when extrapolating, where the model was able to rely upon physical knowledge to overcome a scarcity of measured data. Other novel contributions include the design and undertaking of an experiment, within a laboratory wave tank, to investigate wave loading across a range of representative ocean state spectra. A monopile structure, representing an offshore wind turbine, was heavily instrumented with accelerometers, strain gauges, flow meters, wave gauges and, most importantly, a force collar. The availability of wave load measurements, even within an experimental setting, are rare, making the dataset a valuable resource with which to develop and validate models. Within the thesis, the dataset was used to construct PIML models that relied on only incoming wave height as a model input, a commonly available variable on many offshore structures. Aspects of linear wave theory were integrated within a GP-NARX framework to remove the requirement for access to measured flow conditions close to a structure. The installation of flow meters in offshore environments is typically both expensive and challenging. In the latter stages of the thesis, after a range of PIML models have been developed, a wider view of PIML is taken. Being a relatively new field, the rules of best practise of how to develop models for a given scenario (available level of physical knowledge and measured data) are not yet in place. Relationships between how different types of physics may be incorporated within a model, the effects of changing model structure and the capabilities of constructed models are studied and discussed. The concluding contribution of the thesis is the development of a framework to aid with how best to integrate a given piece of physical knowledge within a PIML model.

Degree

thesis:*
Name dc:type.qualificationname
Ph.D
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
University of Sheffield
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Pitchforth, Daniel James
Advisors dc:contributor.advisor
  • Cross, Elizabeth
  • Rogers, Tim
ORCID person:identifier.orcid

Rights

Language dc:language.dcterms_iso639_2
eng

Identifiers

dc:identifier.*
Dc Identifier Grantnumber
EP/S001565/1

Chain of custody

source
Harvested from
White Rose University Consortium
Base URL
etheses.whiterose.ac.uk/cgi/oai2
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Pitchforth, Daniel James. Physics-informed machine learning for structural dynamics: combining physics and data for offshore structures. doctoral thesis, University of Sheffield, 2025.