From single-task wins to Earth-system understanding: a review charts five barriers and six paths for foundation models in geoscience

AI models can now outforecast weather services and map floods from orbit. Whether that amounts to understanding the Earth system is another question. A review published September 20 in The Innovation Reviews (1: 100014) says the evidence does not yet support the stronger claim — and lays out what would. Earth system science studies a planet in which atmosphere, ocean, land, and ice exchange energy, water, and matter continuously. Data, models, and research communities, however, have long grown up in separation. Foundation models — systems pretrained on broad, multi-source data and then adapted to many tasks — promise to bridge that divide, the authors write, but progress so far is uneven and hard to compare. The review, led by Fenzhen Su and Yinxia Cao of the Chinese Academy of Sciences, sets a bar for what counts as a foundation model: broad pretraining, capability beyond any single task, and evidence of transfer or downstream adaptation. Parameter count and forecast speed alone do not qualify. Comparing six fields — geology, geophysics, atmosphere, ocean, geography, and the integrated Earth system — the authors find that capability tracks data and physics, not ambition. Atmospheric models profit from decades of gridded reanalysis; underground observation is sparse and inversion uncertain; ocean measurements thin out with depth; remote sensing reads the surface well but struggles with coupled human decisions. Standout systems include GeoGalactica, a 30-billion-parameter model pretrained on a 65-billion-token geoscience corpus; a seismic foundation model self-supervised on over 2.2 million unlabeled images; Prithvi-EO-2.0, pretrained on 4.2 million global time-series samples from NASA's Harmonized Landsat–Sentinel-2 record at 30-m resolution; and Aurora, pretrained on more than one million hours of heterogeneous geophysical data and adapted to weather, air quality, ocean waves, and tropical cyclone tracks. The core gap: reusing one model across tasks is not the same as simulating how spheres interact. Coupled modeling demands correct exchange of energy and matter, stable feedbacks, and tolerance to local errors that can amplify — a single accurate component does not guarantee a trustworthy whole. The authors catalog five cross-cutting barriers — an AI-ready Earth data dilemma, fragmented representation spaces, physical inconsistency, deficient scientific validation, and ecosystem barriers around data and computing — and answer with six paths: a unified Earth representation engine, a physics-embedded neural architecture, an Earth system knowledge graph, multi-scale generative inversion, causal reasoning and scenario simulation, and a human-AI collaborative discovery interface. Near term, they recommend data and metadata standards, reproducible benchmarks, physical-consistency diagnostics, and calibrated uncertainty — with models used now for downscaling, bias correction, parameterization, and ensemble generation alongside numerical models. Longer term: coupled generative digital twins, cross-sphere causal models, and autonomous scientific discovery. The paper is a review, not a new model; its paths remain proposals until benchmarks test them. ### Su F., Cao Y., Li X., et al. (2026). Foundation models in Earth system science: Advances, challenges, and future paths. The Innovation Reviews 1(1), 100014. https://doi.org/10.59717/j.tirv.2026.100014