Short-term forecasting of six-degree-of-freedom (6-DoF) motions of a semi-submersible vessel supports the safe execution of motion-restricted offshore operations such as ROV launch and recovery, crew transfers, and crane lifts.
Industry-standard systems reconstruct the surrounding wave field from X-band navigation radar images, propagate it forward using linear wave theory, and convert it to motion through pre-computed Response Amplitude Operators (RAOs).
Existing machine learning approaches to vessel motion forecasting typically rely on past motion alone, sometimes augmented with scalar wave parameters or sparse wave measurements, and therefore do not exploit the full spatial structure of the surrounding wave field.
No prior study has demonstrated direct prediction of 6-DoF vessel motions from raw X-band radar image sequences within a single end-to-end learnable model.

This thesis addresses that gap by proposing the Spatiotemporal Cross-Attention Transformer (SCAT), a neural architecture that takes sequences of X-band radar images together with 45~seconds of past 6-DoF motion and predicts the next 60~seconds of motion.
A convolutional radar encoder produces a tokenised spatiotemporal representation of the radar history, which is queried through cross-attention by learned future motion tokens.
Both radar and motion representations are conditioned on the relative wave--vessel heading through a learned embedding.
Four model configurations are studied by varying the radar input representation and training loss in order to identify the most effective design.
All variants are pre-trained on synthetic data spanning five JONSWAP sea states and eight wave directions, and then fine-tuned on 20~minutes of operational data recorded aboard a semi-submersible rig in the Norwegian Sea.
Their performance is compared against state-of-the-art phase-resolved predictions provided alongside the same recordings and evaluated on four five-hour blocks covering representative wave headings.

On synthetic data, all four variants achieve heave correlations between 0.89 and 0.94 and 6-DoF mean correlations between 0.87 and 0.91 at the 60-second horizon, showing that the architecture can learn wave-to-motion relationships directly from radar imagery.
After fine-tuning, the selected variant reaches heave correlations of 0.69 to 0.76 across all four real-data blocks and reproduces per-DOF directional patterns consistent with the underlying wave physics.
Joint training across all blocks further improves correlation in three of the four blocks.
The state-of-the-art baseline still leads the 6-DoF mean correlation at 60~seconds by 0.14 to 0.31, indicating that synthetic-to-real domain mismatch remains the main limitation.

SCAT is therefore presented as a proof of concept showing that an attention-based model operating on raw radar imagery and motion history can learn physically meaningful 6-DoF motion predictions without hard-coded wave physics, dispersion relations, or RAOs.
These results establish a viable and complementary path toward data-driven maritime motion prediction and suggest that reducing the synthetic-to-real gap through higher-fidelity pre-training data is the most direct route toward operational deployment.

Jellyfish are among the most energy-efficient swimmers in the animal kingdom, driven by unsteady interactions between bell deformation, velar control, and vortex formation. Accurately modelling these mechanisms remains challenging, as existing numerical approaches often rely on simplified geometries, reduced dimensionality, or neglect of velar kinematics, limiting their ability to capture realistic three-dimensional flow dynamics.

This study presents a three-dimensional numerical model of Sarsia tubulosa implemented in the CFD solver WaterLily. The geometry is represented using time-dependent NURBS-based control points, enabling biologically realistic bell deformation and active velar motion. A Cartesian grid combined with the Boundary Data Immersion Method avoids remeshing and allows efficient simulation of complex, time-varying kinematics. Hydrodynamic forces are obtained directly from the incompressible Navier–Stokes equations, avoiding reduced-order approximations.

The model reproduces key propulsion mechanisms, including jet formation and vortex shedding, but discrepancies in force and velocity indicate limitations in physical fidelity. In particular, the absence of fluid–structure interaction, elastic effects, and tentacle-induced drag leads to idealised dynamics and phenomena such as velocity overshoot. This highlights that prescribed kinematics alone are insufficient to fully capture jellyfish propulsion.

Overall, the model provides an efficient framework for analysing three-dimensional flow structures and kinematic trends in jellyfish swimming. However, its use for quantitative efficiency assessment is limited, and future work should focus on incorporating fluid–structure interaction and passive biological effects to improve physical realism.

Code is available in the reproducibility repository that was added as link.:
https://github.com/evanvanderweide/WaterLily-Sarsia-tubulosa-Jellyfish-Modelling

Computational fluid dynamics simulations are commonplace and crucial in many industries, yet obtaining flow solutions requires solving large, sparse linear systems of equations at a significant time and energy cost. This becomes particularly expensive, especially on unstructured meshes. One of the most effective methods for solving such systems is the algebraic multigrid (AMG). Still, its performance depends heavily on the efficiency of its internal components, particularly the smoother. Many existing data-driven AMG enhancements mainly target structured grids or focus on other operators rather than smoothing, leaving a gap in unstructured mesh performance optimization.

This study develops a graph neural network (GNN) inspired by AutoAMG, a GNN-based approach to enhance AMG performance by predicting optimal strength parameters. The GNN in this work is trained to predict polynomial coefficients used to construct a sparse pseudo-inverse smoother matrix, improving the AMG smoothing step. The framework is implemented in PyTorch with graphics processing unit (GPU) acceleration, while the training and evaluation datasets are generated using the CFD solvers WaterLily and ReFRESCO. These cover synthetic problems, canonical external flows, and industry-relevant 2D simulations of airfoils (from the AirfRANS dataset) for both structured and unstructured meshes. The framework includes the graph generation, GNN architecture, AMG integration, and a training and validation pipeline.

Excellent generalization is demonstrated for structured cases, achieving approximately 25% wall-clock time reduction and nearly perfect convergence on grids up to 128 times larger than the ones used in training. Similarly, for unstructured 2D/3D problems, data diversity proves to be the primary driver of generalization. A mixed dataset produces stable acceleration across all meshes, achieving 17-18% solve time reductions on grids up to 20 times larger number of cells. Tests further demonstrate that, while different GNN sizes can enhance solving times, performance is ultimately limited by the smoother itself. Increasing the pseudo-inverse complexity, such as using a high polynomial functions to construct the coefficients, can cause overfitting.

Testing on the complex AirfRANS dataset shows that the model generalizes to significantly larger problems and unseen flow regimes, reducing iterations by nearly 20%. However, due to higher sparse smoother costs, this resulted in a 4.7% net slowdown in solve time. In conclusion, this research demonstrates that a GNN-tuned Jacobi smoother can accelerate AMG solvers on diverse grids, averaging 17-25% wall-clock reductions. The framework demonstrates that a GNN can provide quantifiable speedups, given that the reduction in iteration count outweighs the higher computational cost of the sparse smoother.

The growing demand for sustainable marine transportation has revived interest in Flettner rotors as auxiliary wind-assisted propulsion devices capable of significantly reducing fuel consumption and emissions. While the aerodynamic performance of single rotors has been widely studied, the complex wake-induced interactions between multiple rotors remain insufficiently understood, limiting reliable performance prediction for practical installations. This thesis investigates these wake interactions of tandem finite Flettner rotors through high-fidelity computational fluid dynamics (CFD) simulations supported by data-driven reduced-order modeling (ROM). The study begins with a validation campaign against benchmark experimental data from Shehata and Medina (2021) and numerical data from Liu et al. (2025) for rotating cylinders to ensure the reliability of the employed Biot–Savart bound-ary condition implementation in the WaterLily.jl CFD framework. The validated setup is then extended to perform a systematic parametric study of two interacting rotors by varying spin ratios and incidence angle at a Reynolds number of Re = 1000. The simulations reveal distinct wake regimes governed by up-wash and down-wash interactions, demonstrating how the forces on both rotors change with induced interactions. Additionally, the results indicate that asymmetric spin-ratio combinations can locally enhance the aerodynamic efficiency of the system by mitigating wake interference. To analyze the vast three-dimensional flow database efficiently, Proper Orthogonal Decomposition (POD) is employed. The first few POD modes capture the dominant wake structures responsible for the majority of the flow energy, providing physical insight into the coherent vortical features and their parametric dependence. The POD coefficients serve as compact representations of the flow fields and are used to construct a machine-learning based reduced order model. A multi-layer perceptron (MLP) and a Bayesian regression (BR) model is trained to learn the nonlinear mapping between the global parameters (spin ratio and angle of incidence) and the corresponding POD coefficients. The resulting models are capable of predicting the time-averaged velocity and pressure fields for unseen parameter combinations with reasonable accuracy and low computational cost. The combined POD-ML framework demonstrates the feasibility of predicting complex wake-interaction behavior using limited high-fidelity data. While the models successfully reproduce dominant flow patterns and aerodynamic trends, their accuracy can be further improved by incorporating additional high Reynolds number data and broader parameter coverage. The thesis concludes that reduced-order modeling based on POD and machine learning provides a promising pathway toward fast, data-driven performance prediction tools for wind-assisted propulsion systems featuring multiple interacting Flettner rotors.