Accelerating Earth System Modeling by Leveraging a Standardized Coupling Framework for Hybrid Models

Traditional Earth System Models (ESMs), while foundational to climate and weather research and applications, often face significant computational bottlenecks and struggle with accurately parameterizing complex sub-grid physical processes such as atmospheric convection clouds and radiation. Recent progress in Machine Learning (ML) and Artificial Intelligence (AI) provides a promising solution to these constraints by developing computationally efficient data-driven surrogates for particular model components such as atmospheric solvers, sub-grid scale parameterizations, and data assimilation to reduce biases. This study presents a novel coupling architecture designed to seamlessly couple diverse ML/AI-based model components with traditional physics-based Earth System Models to construct hybrid earth system models. The proposed infrastructure and newly developed GeoGate co-processing component standardizes the communication and interactions among the different types of model components, enabling asynchronous execution and flexible swapping of traditional physics components with their AI/ML-based data-driven counterparts to establish a true hybrid modeling environment. We demonstrate that this generic coupling framework allows adaptation to different ESM applications, such as regional configuration that couples atmosphere-ocean to more sophisticated and multi-component applications that aim for Subseasonal-to-Seasonal (S2S) type predictions. The proposed framework enhances model development flexibility, paving the way for the creation of next-generation computationally efficient and modular ESMs that can leverage the rapid innovation cycles of both physical and machine learning sciences in Earth system modeling while maintaining and, in some cases, improving simulation fidelity across various spatial and temporal scales.