Bridging Climate Projections and Coastal Physics: Exploring Generative AI for High-Temporal Wind Downscaling

Assessing the impact of climate change on critical coastal infrastructure requires transition from global climate signals to local hydrodynamic responses. While Global Climate Models (GCMs) provide necessary long-term projections, their coarse resolution fails to resolve the localized wind gradients and fetch-limited dynamics essential for modeling extreme 10,000-year storm surges. The fidelity of the meteorological forcing—specifically at high temporal resolutions—is the primary bottleneck for reliable surge prediction.

This study investigates a downscaling workflow that prioritizes the coupling between climate data and DHI’s high-resolution coastal ocean models. Leveraging a library of existing metocean datasets from regional numerical weather models, we present a comparative study aimed at achieving a target resolution of 3 km and, crucially, 10 minutes. This high temporal frequency is vital for capturing the non-linear energy transfer and peak wind stresses that drive extreme water levels in the Danish Straits.

We establish a tiered benchmarking framework to evaluate downscaling performance across meteorological and hydrodynamic indicators. Classical interpolation methods serve as the baseline to quantify the added value of more sophisticated approaches. We then explore non-linear generative models, such as conditional Generative Adversarial Network (cGAN) architecture, against an approach utilizing Stochastic Interpolants (SIs). The latter is examined for its potential to better preserve the kinetic energy spectrum and turbulent trajectories of wind fields, ensuring that the generated 10-minute forcing remains physically consistent with the requirements of numerical solvers like MIKE 21/3.

Evaluation focuses on the operational utility of the downscaled forcing for numerical solvers like MIKE 21/3. We demonstrate how the chosen AI methodologies recover the non-linear "tails" of extreme event distributions which are typically smoothed out by traditional interpolation.