Pathways to Coupled Ocean-Atmosphere AI/ML Forecasting: From Medium Range to Subseasonal Scales

A high-resolution global data-driven coupled system has high potential to deliver enhanced skill for extreme events on longer forecast horizons (5-90 days). The convergence in new techniques, higher resolution models and data, and incorporation of the ocean’s role in weather dynamics on longer time spans make this a prime moment to achieve significant advances, while also recognizing the limitations and learnings from past and recent AI/ML models. Given the rapid and demonstrative success of diffusion-transformer based approaches to video-generation models such as OpenAI’s latest release, sora, we wish to investigate whether the architectural and training design decisions, which give rise to their apparent emerging capability to preserve 3D consistency, long-range temporal coherence, and object permanence, can be similarly leveraged to produce next-generation weather foundation models. Alongside we will explore whether emerging hybrid physics-data models have utility for the earth system components. That is, is a neural ESM a potentially viable means of advancing probabilistic forecasting on longer time horizons? We will identify a case study to anchor the feasibility assessment of the proposed AI/ML transformer-diffusion system. Possible high impact weather events to consider include an intensifying tropical cyclone (e.g., Super Typhoon Mawar in the western Pacific in May 2023), a severe storm in the North Atlantic (e.g., Storm Isha in Jan 2024 impacting the UK), or other events with outlier attributes. Parameters such as rainfall rate, peak winds, and peak near-surface air temperature will be assessed along with CRPS and Brier Skill Score. We will aim to develop two versions of the provisional foundation model, an atmosphere-only and a coupled ocean-atmosphere model. Lastly, within a coupled model framework, we will consider the sensitivity of results to the resolution of the ocean and atmosphere training data. We will then try the configurations we have discussed above, including atmosphere-only/coupled, resolutions, input data resolutions, for the case study - first at a short lead time, then increasing to longer lead times in the Option period.