An AI weather model learns what a storm does to the sea

Hurricanes Idalia and Franklin leave a trail across the sea: higher waves and a cold wake of stirred-up water. In AIFS Marine’s research forecasts, begun on 26 August 2023, those changes develop alongside the storm.[1]

The European Centre for Medium-Range Weather Forecasts (ECMWF) highlighted those reruns on 12 October while explaining AIFS Marine, a prototype combining atmosphere, waves, sea ice and surface ocean. Researchers report roughly one extra day of marine forecast skill against ECMWF’s physics-based systems for almost all assessed variables at medium range, in deterministic forecasts that give a single predicted outcome. It corresponds to comparable average error at a later forecast time, with the gain varying by quantity.[1]

The useful part is the connection

A list of ocean predictions can sound like an expanded weather menu. The deeper change is that the ingredients evolve together. Wind builds waves; ice damps them; ocean temperatures influence the air above. A forecast can get individual quantities approximately right while making their combination implausible.[3][4]

Conventional forecasting systems, including ECMWF’s Integrated Forecasting System (IFS), exchange information between specialised atmosphere, ocean, wave and ice models. AIFS Marine puts their fields inside one learned system, training on the combined state as it changes.[2]

The historical record makes that possible. Ocean measurements are patchy, so ECMWF’s ORAS6 reanalysis combines observations with models governed by physical laws to reconstruct past conditions. Its improved treatment of surface temperature captures daily warming and cooling more closely against drifting observations. That supplies the kind of changing ocean history an AI model can learn from.[3]

Physical modelling therefore remains part of the chain of knowledge. Observations constrain a reconstruction; the reconstruction becomes training material; a learned forecast attempts to carry those relationships forward. Better ocean observations and better reconstructions can benefit both conventional and AI forecasting.[3]

What the comparisons actually measured

Sea-surface height was the main exception, developing an increasingly negative bias with forecast length. The authors suspect the historical sea-level trend contributes.[2]

The ocean already entering the forecast

There is an operational foothold for this work. On 12 May, ECMWF put AIFS version 2 into service with wave components in both its Single and Ensemble systems, making 11 wave-related variables available.[4]

ECMWF’s physical forecasting system improved at the same time, adding its NEMO4-SI³ ocean and ice model and more detailed wave–ice and wave–current interactions. Both approaches are being developed around the same practical requirement: a useful marine forecast must describe the way its parts affect one another.[4]

The next research steps reach towards probabilistic marine forecasts and deeper ocean layers. The hurricane’s cold wake captures their purpose: the atmosphere leaves a changed sea behind it, and that changed sea belongs in the next forecast.[1]

Report an error or suggest a correction