LangYa paper: a single AI model now forecasts the global ocean
A Chinese research team just published the blueprint for one of the more ambitious AI-for-science systems in the water: a single neural network that forecasts the state of the entire global ocean up to a week out.
The Chinese Academy of Sciences' Institute of Oceanology has published the peer-reviewed framework behind LangYa, a large AI model that generates one-to-seven-day forecasts of the global ocean from a single network — no recursive stepping, no ensembles of specialist sub-models.
The paper, out in Science Bulletin this month, describes a system that predicts temperature, salinity, and three-dimensional currents across 32 depth layers — 128 variables in all — at a 1/12-degree resolution, roughly nine kilometers per grid cell at the equator. The headline innovation is architectural: LangYa uses a large-language-model-style time embedding to encode the forecast lead time explicitly, so the model can jump straight to a five-day forecast without rolling through days one through four the way earlier AI ocean systems do. That removes a major source of error accumulation, and it lets one network do work that previously required several models stitched together.
The team trained LangYa on 27 years of GLORYS12 ocean reanalysis data (1993–2019) paired with ERA5 atmospheric forcing, holding out 2020–2021 for testing. On seven-day global forecasts, the reported root-mean-square errors came in under 0.0736 m/s for zonal currents, 0.0701 m/s for meridional currents, 0.4376 °C for temperature, and 0.1302 psu (practical salinity units) for salinity. The authors say those numbers compare favorably with existing open-source AI ocean forecasters and independent observations, particularly on multi-variable consistency, lead-time stability, and the thermocline — the layer between tens and roughly 300 meters down where temperature changes fastest. To capture that layer, LangYa adds a thermocline-focused adaptive loss function, a cosine-attention ocean module that stabilizes training on high-dimensional global data, and an asynchronous cross-iterative sampling strategy to represent how the atmosphere actually drives ocean change.
The context matters: LangYa is a family, not a debut. Version 1.0 launched in late 2024 and is already deployed at China's National Marine Environmental Forecasting Center; version 2.0, released in June, added six specialized vertical models for typhoons, storm surges, internal waves, mesoscale eddies, precipitation, and sea ice — including a three-kilometer Arctic sea-ice forecaster built to support navigation on the Northern Sea Route. The new paper is the scientific foundation under that system, and it signals where China's AI-for-science push is heading: unified foundation models replacing brittle pipelines of specialist components, with specialized phenomenon models bolted on top. The work was led by associate researcher Nan Yang, with Fan Wang and Xiaofeng Li as corresponding authors.
What to watch: the authors say the current version is limited to short-term forecasting, with longer lead times and phenomenon-level prediction flagged as the next milestones. The real test of a paper like this is operational — whether LangYa's single-model approach holds up in continuous real-time use, where numerical models with four decades of tuning remain the default.
Would you trust an AI ocean forecast for shipping or storm evacuation over the numerical models agencies have used for decades? Tell us in the comments.
