Sampling sea state using a diffusion model

Research Question and Approach

Traditional spectral wave models, such as WAVEWATCH-III, are computationally expensive, making them difficult to use for ensemble-based probabilistic forecasting or online coupling in climate models. While deep learning has improved weather forecasting, existing AI-based wave models are often deterministic and limited to bulk variables like significant wave height. This study introduces a denoising diffusion probabilistic model (DDPM) to sample the complex conditional distribution of global sea state variables. By conditioning on a 5-day history of global wind forcing, the model directly generates sea state snapshots without the need for autoregressive time-stepping.

The diffusion model successfully captures both bulk variables (significant wave height, mean period, and direction) and more complex partition-related and derived variables (such as Stokes drift and mean square slope). The model achieves substantial computational acceleration compared to numerical spectral models while maintaining high skill and providing a well-calibrated ensemble spread for bulk variables. The authors demonstrate that the model effectively identifies regions of multi-modal sea states, such as crossing seas, and preserves the physical consistency of wind-wave growth relationships, such as Toba's law.

This approach offers a scalable alternative for probabilistic wave forecasting, which is critical for operational maritime applications and for integrating wave dynamics into broader Earth system models. By moving beyond deterministic predictions, the model provides valuable information about forecast uncertainty and enables the estimation of variables that are otherwise too costly to compute, facilitating better parameterization of air-sea interactions.