Bridging Oceans and Atmosphere: A More Comprehensive Weather Model

ICLR 2026 Conference Submission15741 Authors

Keywords: Ocean, Atmosphere, Coupler

Abstract: The Earth system synthesizes multivariate interactions across atmospheric processes, oceanic circulations, cryospheric dynamics, and radiative forcing. While machine learning has transformed weather prediction within individual domains, current models inadequately capture cross-component couplings critical for holistic Earth system modeling. Fundamental challenges emerge from the disparities in spatial and temporal scales and prohibitive computational costs of training integrated models ab initio. To overcome these limitations, we present the Coupled Ocean-Atmosphere Framework (COAF), a deep learning architecture that dynamically couples pre-trained atmospheric and oceanic models through adaptive spectral transformations designed to resolve scale mismatches. COAF introduces two pivotal innovations: An effective structure enabling energy-momentum exchange across domains without structural overhauls of existing models, and an online replay buffer mechanism that enhances long-term stability. Experimental results demonstrate COAF's effectiveness in operational scenarios, achieving a 10\% reduction in latitude-weighted RMSE for key prognostic variables (Z500, T2m) beyond 10-day forecast horizons. These advancements establish a new paradigm for coupled Earth system modeling that balances physical consistency.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15741