Technical Summary: Earth Science Foundation Models: From Perception to Reasoning and Discovery

Problem Statement

Earth science has evolved from independent subdisciplines into an integrated study of the coupled Earth system (atmosphere, hydrosphere, lithosphere, biosphere, anthroposphere, and cryosphere). While the Scientific Data Collection stage has yielded massive, heterogeneous multimodal datasets (seismic records, satellite imagery, gridded reanalysis, and domain text), conventional analytical workflows struggle with their scale, heterogeneity, and cross-sphere complexity.

Existing Artificial Intelligence (AI) applications in Earth science have largely been confined to Scientific Perception—task-specific models for classification, segmentation, or feature extraction within isolated domains (e.g., optical remote sensing or specific weather forecasting). These approaches often fail to address core geoscientific problems that require integrating heterogeneous observations across modalities and scales, reasoning over coupled physical processes, and supporting multi-step analytical workflows. Furthermore, current literature lacks a systematic analysis of the developmental trajectory of Earth Foundation Models (Earth FMs) from isolated predictive systems to dynamic, tool-augmented agents capable of Scientific Discovery.

Methodology and Framework

This paper provides a unified review of Earth Science Foundation Models (Earth FMs) through two complementary dimensions: Depth and Breadth.

1. Depth (Capability Evolution): The authors trace the evolution of model capabilities through three distinct stages:

   • Scientific Perception: Task-specific models (e.g., CNNs, Vision Transformers) optimized for fixed objectives like land-cover classification or change detection.
   • Scientific Reasoning: The emergence of large foundation models (LLMs, multimodal transformers) designed to learn transferable representations from large-scale geospatial and scientific data. These models support cross-modal understanding, complex scientific reasoning, and zero-shot transfer.
   • Scientific Discovery: The transition toward autonomous or semi-autonomous agentic systems. These systems can plan workflows, invoke external scientific tools, interact with simulations, and iteratively refine outputs to support end-to-end scientific investigation.

2. Breadth (Application Scope): The review categorizes existing Earth FMs across the six major Earth spheres and their coupled processes. The paper systematically analyzes domain-specific models for the Atmosphere, Hydrosphere, Lithosphere, Biosphere, Anthroposphere, and Cryosphere, as well as emerging unified models that integrate these spheres.

3. Data Compilation: To support this review, the authors compiled a curated list of over 200 datasets and benchmarks spanning diverse Earth science tasks and modalities, including textual corpora, structured records, visual imagery, and time-series grids.

Key Contributions

• Unified Taxonomy: The paper establishes a structured taxonomy for Earth AI, moving beyond modality-centric views (e.g., remote sensing only) to cover computational challenges across all six Earth spheres. It explicitly maps the shift from perception to agentic discovery.
• Comprehensive Survey of Earth FMs: The authors review representative multimodal Earth foundation models, detailing their architectures (e.g., Vision Transformers, Diffusion models, Graph Neural Networks, and Multimodal LLMs) and their specific applications in tasks ranging from weather forecasting to mineral exploration.
• Resource Curation: A significant contribution is the compilation of a repository containing over 200 datasets and benchmarks, accessible via a maintained GitHub list, to support community efforts in developing Earth FMs.
• Roadmap for Agentic Workflows: The paper details the transition from static prediction to Agentic Scientific Workflows. It reviews emerging frameworks (e.g., PANGAEA GPT, Earth-Agent) that utilize tool-augmented agents for hypothesis generation, experiment execution, and autonomous analysis.
• Identification of Challenges: The paper outlines critical barriers to the next generation of Earth AI, including:
  • Data Heterogeneity: The difficulty of constructing unified Earth embeddings across disparate modalities and scales.
  • Scientific Reliability: The need for continual updating, physical consistency, and trustworthiness in high-stakes scenarios.
  • Scalability and Sustainability: The computational and environmental costs of training large models on high-dimensional Earth data.
  • Embodied Intelligence: The challenge of moving from digital tool-use to agents that can actively acquire data and interact with physical environments (e.g., underwater drones).

Results and Findings

The paper synthesizes the current state of the field, highlighting several key trends:

• Domain-Specific Maturity: Significant progress has been made in domain-specific FMs. For example, in the Atmosphere, models like PanGu-Weather and Aurora have achieved operational-level forecasting speeds and accuracy. In the Lithosphere, models like SeisT and MineAgent are enhancing earthquake monitoring and mineral exploration through multimodal integration.
• Rise of Unified Models: Unified Earth FMs (e.g., Copernicus-FM, TerraMind, PANGAEA GPT) are emerging to bridge silos, demonstrating improved performance in cross-sphere tasks (e.g., linking land cover changes to atmospheric quality) compared to single-sphere models.
• Agentic Capabilities: Early agentic systems (e.g., ClimateAgent, OpenEarthAgent) are demonstrating the ability to orchestrate complex workflows, such as querying databases, running simulations, and generating code, effectively acting as "AI co-scientists."
• Reasoning Enhancements: The paper notes a shift from simple supervised fine-tuning to Reinforcement Learning with Verifiable Rewards (RLVR) to improve the logical deduction and "faithful reasoning" capabilities of geospatial models, addressing issues of shallow reasoning and hallucination.

Significance and Future Outlook

The paper posits that Earth AI is undergoing a paradigm shift from isolated predictive tools to integrated frameworks capable of addressing macro-scale challenges like global climate change and natural disaster management autonomously.

The significance of this work lies in its provision of a structured roadmap for the development of Earth Foundation Models. By defining the trajectory from perception to reasoning and discovery, the authors aim to guide the community toward creating integrated, trustworthy, and actionable AI Earth Scientists.

The paper concludes that the ultimate frontier is the convergence of unified Earth FMs, skill-based agentic reasoning, and embodied spatial intelligence. This vision involves systems that not only analyze existing data but also actively seek new evidence, coordinate measurements across platforms, and close the loop between scientific hypothesis, observation, and action. Achieving this requires close collaboration across geoscience, remote sensing, machine learning, and robotics to realize a new generation of autonomous AI Earth Scientists capable of supporting end-to-end scientific discovery and planetary-scale stewardship.