Toward Artificial Intelligence–Enabled Earth System Coupling

Maria Kaselimi, Anna Belehaki · 2026

Read Toward Artificial Intelligence–Enabled Earth System Coupling by Maria Kaselimi, Anna Belehaki (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-03289.txt.

Introduction

In their 2026 review, Kaselimi and Belehaki ask how AI methods can tighten links across Earth system parts. Rather than survey new models broadly, they focus on links between domains and on joined multi-part pictures. They list chances, lasting limits, and ideas for unified frameworks with better physical consistency and models people can read.

Catalog record

CREATORMaria Kaselimi, Anna Belehaki
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73116
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.03289

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Toward Artificial Intelligence–Enabled Earth System Coupling Maria Kaselimi, Anna Belehaki Abstract Coupling constitutes a foundational mechanism in the Earth system, regulating the interconnected physical, chemical, and biological processes that link its spheres. This review examines how emerging artificial intelligence (AI) methods create new opportunities to enhance Earth system coupling and address long-standing limitations in multi-component models. Rather than surveying next-generation modelling efforts broadly, we focus specifically on how state-of-the-art AI techniques can strengthen cross-domain interactions, support more coherent multi-component representations, and enable progress toward unified Earth system frameworks. The scope extends beyond climate models to include any modelling system in which Earth spheres interact. We outline emerging opportunities, persistent limitations, and conceptual pathways through which AI may enhance physical consistency, interpretability, and integration across domains. In doing so, this review provides a structured foundation for understanding the role of AI in advancing coupled Earth system modelling. keywords Earth system, interactions, coupling, physics-informed machine learning, graph neural networks, causal discovery 1 Introduction The Earth system is commonly described as comprising four major, dynamically interacting components—the atmosphere, hydrosphere, geosphere, and biosphere. In many frameworks, these are complemented by additional subsystems, notably the cryosphere (snow, sea ice, glaciers, and permafrost) and the anthroposphere, which encompasses human activities and socio-environmental processes. These spheres do not operate in isolation; rather, they are tightly coupled through the continuous exchange of energy, mass, momentum, and chemical species. Such exchanges drive variability and feedbacks across a wide range of spatial and temporal scales, shaping phenomena from weather and hydrology to climate dynamics and biogeochemical cycles. Within this context, coupling refers to the explicit representation of these bidirectional or multilateral interactions in numerical models and data-driven frameworks. Coupling captures the mechanisms through which perturbations in one subsystem propagate to others, altering their state and behavior and ultimately modulating the evolution of the Earth system as a whole 32 ; 64 . Traditional earth modelling rely on a modular coupling architecture in which separately developed component models—such as the atmosphere, ocean, land surface, and cryosphere—are interconnected through a central coupler 64 . This design has enabled major advances in climate and weather prediction, but several persistent limitations remain. High-resolution coupled simulations are computationally expensive, restricting ensemble size and limiting systematic exploration of uncertainty. Structural and parametric errors originating in one component can propagate across interfaces, producing long-term drifts in energy, water, or carbon budgets 4 . Scale mismatches between resolved dynamics and parameterized subgrid processes complicate the representation of interfacial fluxes 40 , while heterogeneous spatial grids, remapping algorithms, and asynchronous time stepping introduce additional numerical inconsistencies 100 . The joint assimilation of multi-sphere observations—necessary to maintain physically consistent coupled states—further increases methodological complexity and computational cost 68 ; 19 . These limitations underscore the need for complementary approaches capable of representing coupling processes more flexibly while preserving physical integrity. Recent advances in artificial intelligence (AI) and machine learning (ML) have opened new avenues for improving Earth system prediction, particularly through the emergence of large-scale foundation models. Early Earth-observation and atmospheric foundation models demonstrated impressive capabilities in learning coherent, multi-variable representations from vast and heterogeneous datasets. Examples include NASA–IBM’s Prithvi-EO-2.0 83 , Microsoft’s TerraFM 25 , and Google DeepMind’s TerraMind 44 , which provide powerful multimodal encodings of satellite observations, as well as atmospheric models such as ClimaX 61 , which learns variable-agnostic representations of reanalysis fields and supports a wide range of downstream tasks. These systems demonstrated that foundation models can extract transferable structure from large geoscientific datasets and enable flexible generalization across variables, sensors, and regions. Building on these advances, recent efforts have begun to extend foundation models beyond single domains toward architectures that can directly represent multi-sphere interactions. Coupling-oriented foundation models—including Microsoft’s Aurora 8 , European Centre for Medium-Range Weather Forecasts (ECMWF) AI Forecasting System (AIFS) 51 , and ORBIT 95 —are trained on integrated atmospheric, land, ocean-wave, and climate-mode datasets, enabling them to capture cross-component dependencies and teleconnection patterns within a unified framework. Their distinguishing feature is the ability to approximate aspects of coupling. Importantly, the value of these models arises not from AI alone but from the combination of large-scale learning with embedded domain knowledge, including physical constraints, curated multi-sphere training corpora, and architectures designed to respect geophysical structure 10 . This integration makes coupling-oriented foundation models a promising complementary pathway for representing the physically mediated exchanges that govern the behavior of the atmosphere, ocean, land, cryosphere, and human systems. At the same time, a growing body of research is moving beyond purely data-driven strategies toward AI approaches that explicitly integrate domain knowledge, physical structure, and meaningful interactions among system components. Physics-informed ML and hybrid modelling frameworks embed conservation laws and process-based constraints, ensuring that learned representations remain consistent with established scientific principles. Neural operators learn continuous dynamical mappings and can be combined into multi-component surrogates that respect underlying functional relationships. Graph neural networks (GNNs) provide a natural way to encode spatial, temporal, and cross-variable interactions on irregular meshes and heterogeneous domains, while causal discovery and explainable AI offer complementary tools for identifying directional influences and revealing the structure of complex system behaviour. These emerging directions reflect an important shift in AI research—from treating Earth system prediction as a pattern-recognition problem to designing architectures that capture mechanisms, dependencies, and scientifically grounded reasoning 90 . The purpose of this review is to synthesize and organize these emerging developments with explicit focus on coupling. Rather than surveying AI in Earth sciences broadly, we concentrate on methods whose primary role is to represent, emulate, constrain, or interpret interactions across components of the Earth system. Building on the physical perspective of coupling introduced in Section 3 and the description of traditional coupling frameworks in Section 4 , then in Section 5 we examine how AI can be integrated into these infrastructures and what new risks and opportunities this integration entails. Figure 1: Earth System Interaction Diagram Specifically, the objectives of this review are to: • Characterize the role of coupling in Earth system modelling and delineate the numerical and conceptual limitations of traditional coupling frameworks. • Provide a detailed exploration on AI approaches for Earth system coupling, including interface and flux emulators, neural-operator-based surrogates, graph-based interaction models, multi-domain foundation models, and causal/interpretability methods. • Assess the benefits and limitations of these AI approaches in coupling contexts, with emphasis on physical consistency, long-term stability, predictive skill, computational efficiency, and uncertainty quantification. • Highlight the role of multi-component observational datasets—including satellite, in-situ, reanalysis, and socioeconomic data—in enabling AI-based coupling solutions and multi-domain data fusion. • Outline evaluation strategies and open research directions for integrating AI-based couplers into next-generation ESMs and digital twin frameworks. 2 Scope and Boundaries of This Review This review does not aim to provide a comprehensive assessment of next-generation Earth System Models (ESM) in their entirety, nor does it attempt to catalogue the full landscape of advances in climate modelling. Instead, its scope is intentionally focused on coupling—that is, on the physical and computational mechanisms that mediate interactions across the atmosphere, hydrosphere, geosphere, biosphere, and cryosphere. By design, this review concentrates on how emerging AI approaches can strengthen these interactions, address long-standing numerical limitations, and offer conceptual pathways toward more unified and physically coherent multi-component models. Because our

Read the complete text