GNN-Autoencoders as Surrogates for Coastal Ocean Modeling with MIKE 21 HD

Type of project

M.Sc.

Contact

Jesper Mariegaard (jem@dhigroup.com), Rocco Palmitessa (rpal@dhigroup.com)

Topic

Graph Neural Networks (GNNs) have shown great potential in learning from structured spatial data, such as water levels and currents on unstructured meshes. In this project, we propose to investigate whether a GNN-based autoencoder can be trained to emulate MIKE 21 HD simulations. The idea is to learn a latent representation of the dynamics and use it for fast surrogate predictions. Two real-world test cases, the Southern North Sea and the Øresund, will be used for model training and validation.

Relevance to DHI

Physics-based models like MIKE 21 HD are computationally expensive, especially in large or ensemble-based setups. Fast surrogate models, such as GNN-based autoencoders, could enable new use cases such as long climate scenarios, real-time predictions, probabilistic modeling, or data assimilation – at a fraction of the computational cost.

What is the student's contribution

Implement and train a GNN-autoencoder on MIKE 21 HD results for the selected test cases. Evaluate reconstruction quality and runtime speedup. Compare performance and interpretability to an existing simpler surrogate framework FM-ROM. Deliverables: code, trained models, and feasibility assessment.