Koopman Autoencoders as surrogates for MIKE 21 coastal models with integration of observed data

Type of project

M.Sc.

Contact

Jesper Mariegaard (jem@dhigroup.com), Freja Høgholm Petersen (frtp@dhigroup.com)

Topic

Koopman autoencoders are deep learning models that learn a linear latent space where nonlinear dynamical systems can be evolved using linear dynamics. This project explores implementing a Koopman autoencoder with external forcing, inspired by Morton et al. (2018), and applying it to MIKE 21 HD simulations in the Øresund and Southern North Sea. The student will investigate how observations (e.g., water levels) can be integrated during training to improve predictions. Additional focus areas may include multi-step forecasting, stability analysis, or formulation as a gResNet-type model.

Relevance to DHI

MIKE 21 HD models are accurate but computationally expensive. Learning surrogate models in Koopman space could enable long climate scenarios, fast forecasting and efficient data assimilation, combining physics-based knowledge with observational data. This aligns with DHI’s long-term goals of hybrid AI-physics modeling.

What is the student's contribution

Implement a Koopman autoencoder (with dynamic inputs) and apply it to one or two test cases. Compare performance and interpretability to an existing simpler surrogate framework FM-ROM. Experiment with integrating observations in the training process. Deliverables: code, test results, and feasibility assessment.