Koopman Autoencoders Outperform Traditional Methods in Coastal Ocean Modeling
A recent study presents a versatile formulation of the Koopman autoencoder for modeling coastal-ocean dynamics, integrating meteorological influences and boundary conditions. This research rigorously contrasts this method with conventional proper orthogonal decomposition (POD)-based surrogates, which have been extensively utilized in hydrodynamic contexts. The Koopman autoencoder features a learned linear temporal operator in latent space, enhanced by eigenvalue regularization to ensure temporal stability. Additionally, this approach is assessed with temporal unrolling methods for reliable long-term forecasts. The evaluation included three distinct test cases, with prediction periods reaching one year at a 30-minute resolution. The findings, documented as arXiv:2602.05416v2, highlight the effectiveness of reduced-order surrogates with temporal unrolling in achieving significant accuracy.
Key facts
- The paper introduces a flexible Koopman autoencoder formulation for coastal-ocean modeling
- The approach incorporates meteorological forcings and boundary conditions
- Performance is systematically compared against POD-based surrogates
- The Koopman autoencoder uses a learned linear temporal operator in latent space
- Eigenvalue regularization promotes temporal stability
- Temporal unrolling techniques are evaluated for stable long-term predictions
- Models were tested on three cases with distinct dynamical regimes
- Prediction horizons extended up to one year at 30-minute resolution
Entities
—