Foundation Models of Ocean and Atmosphere in 2025: Milestones and Perspectives

Abstract

Foundation models–large neural networks pretrained on heterogeneous Earth-system datasets and adapted to multiple tasks–are reshaping atmospheric and oceanic prediction. In this overview, we contrast this paradigm with task-specific supervised learning, review major atmospheric models (FourCastNet, ClimaX, Pangu-Weather, GraphCast, FengWu, FuXi, Prithvi WxC, AIFS, Aurora) and the first wave of oceanic models (WenHai, XiHe, LangYa), outline applications (forecasting, downscaling, statistical correction, subseasonal/seasonal prediction, analogs discovery, climate risk), and discuss challenges and perspectives including upcoming Russian initiatives.