Foundation models in Earth system science: advances, challenges, and future paths

Abstract

The convergence of massive Earth observation datasets and emerging foundation models offers an opportunity to address the complex challenges in Earth System Science (ESS). However, the direct application of foundation models to ESS is often hindered by domain gaps, characterized by the need for physical consistency, the complexity of Earth’s spherical geometries, and the non-stationarity of the multisphere coupled Earth systems. Building on prior reviews that prioritize model architecture or view geoscience as a monolithic entity, this study provides a domain-centric synthesis of foundation models across six distinct subfields, including Geology, Geophysics, Atmosphere, Ocean, Geography, and the Integrated Earth System. By analyzing the applications and limitations of Artificial Intelligence (AI) within these sub-fields, this work identifies unique domain barriers and cross-cutting challenges, including the AI-ready Earth data dilemma, representation space fragmentation, physical inconsistency, scientific validation deficiency, and ecosystem barriers. In this context, we provide future paths for the next generation of ESS foundation models, driven by unified Earth representation engine, a physics-embedded neural architecture, an Earth system knowledge graph, multi-scale generative inversion, causal reasoning, and human-AI collaborative discovery interface. In this way, foundation models can potentially complement traditional numerical models to monitor planetary status and simulate its future evolution, thereby equipping the scientific community with new tools for holistic planetary stewardship and complex decision-making.