This thesis investigates the application of data-driven surrogate models within a multi-fidelity framework to aid and accelerate the structural design of motion-compensated offshore gangways. The motivation was found in the literature review, where gaps were found regarding suitability for higher-dimensional problems and multi-type input, sampling techniques, suitability for data-scarcity and suitability for application on complex mechanical systems.
The literature review starts by establishing the foundations of surrogate modelling, including its evolution from response surface methods and Kriging, to more advanced data-driven techniques such as support vector regression and neural networks. A distinction is made between physics-based, data-driven, and hybrid models, highlighting trade-offs in terms of complexity, interpretability, and data requirements. Challenges such as generalizability, extrapolation, and the curse of dimensionality are addressed, alongside the importance of design of experiments (DoE) in constructing effective models. Classical and modern DoE methods are discussed, including Latin hypercube sampling and space-filling strategies, as well as adaptive sampling techniques. Recent trends in the field include hybrid models that integrate physical laws into machine learning algorithms, and multi-fidelity frameworks that fuse low- and high-fidelity data to reduce computational cost. These advances are framed as promising solutions to limitations in data efficiency and model applicability to real-world engineering problems. The case study centers on an offshore gangway structure designed by Ampelmann Operations B.V., a company specializing in motion-compensated systems for transferring personnel and cargo to offshore platforms. The gangway operates under diverse loading conditions such as live loads, environmental forces, and cargo transfers, resulting in thousands of unique load combinations. These are categorized according to certification standards and reduced to 4016 discrete cases for model input. Structural data from an existing, fully designed system forms the basis for both model training and evaluation. A detailed description of the gangway, including its main components and loading scenarios, is provided
to contextualize the design space. The methodology involves several key stages. First, the data of the case study is described in detail. This involves defining a numerical representation of structural severity to act as the model output, condensing global responses such as stress, deflection, and buckling into a single ranking metric for each load case. A low-fidelity analytical model is implemented in Python to provide fast approximations of structural response, extended to include stress and buckling calculations. A high-fidelity finite element
model is provided by the partner company, which is simplified to ensure numerical stability. As an alternative, a simple truss model is used for better suitability for beam model analysis in FEM. As a preliminary sampling, random sampling is proposed for the design of experiment. Then, an intelligent sampling technique based on Euclidean distances is proposed to improve data efficiency by selecting high-fidelity samples in underrepresented regions. After data sampling, a machine learning framework is created to support data-driven surrogate modelling. Three model types are implemented: Ridge regression, KRR and XGBoost. These were implemented after pre-processing the data, and a random grid search is implemented for hyperparameter tuning, combined with cross-validation. The selected performance metrics are the REP, MAPE, PPMCC and the R-squared score. A multi-fidelity framework is introduced to combine predictions from low- and high-fidelity sources using the comprehensive approach.
The results show that modelling issues in the high-fidelity model lead to data that is not physically plausible. Among the surrogate types tested, Ridge regression is inadequate for complex, mixed-type input features. KRR performs moderately, and XGBoost demonstrates strong performance particularly in scenarios with limited high-fidelity data. The intelligent sampling strategy does not improve the performance compared to random sampling. The multi-fidelity model yields marginal improvements for KRR on the complex model, but it does not achieve better performance then a single-fidelity XGBoost model. The simple truss model showed an improvement of XGBoost performance when data is scarce, which confirms the hypothesis. As the simple truss model can be extended to cover the full range of load cases, this could justify the time required for implementing the multi-fidelity framework. However, in
the current state the benefit of multi-fidelity is marginal. Regarding the complex dataset with modelling issues in the source, it is still useful that this research found a well-predicting surrogate in XGBoost, even though multi-fidelity might not be the answer for that dataset due to badly correlated data. The predicting surrogate, proven to have well-predictive capabilities even though the data represents modelling issues, can point engineers to all load cases that result in exceeding members. Therefore, the surrogate could aid engineers by acting as a pointer to load cases that require engineering judgment. Also, the predictive capabilities of the surrogate enable surrogate-based optimization. Both findings can increase efficiency of the structural design of the gangway. The study concludes that surrogate modelling, when carefully implemented, can improve the efficiency of structural design evaluations for offshore gangways. XGBoost appeared to be best suited for this case involving complex, moderate-dimensional input spaces. Multi-fidelity modelling is viable but sensitive to data quality. Future work should focus on extending the framework to surrogate-based optimization and exploring adaptability to different geometries