Artificial intelligence to support cross-disciplinary climate change research

Professor Haewon McJeon’s team at KAIST, together with Prof. S. Karthik Mukkavilli and Prof. Alice Oh and an international research team from Peking University, Imperial College London, Politecnico di Milano, the University of Maryland, and the International Institute for Applied Systems Analysis (IIASA), published a commentary for an AI foundation-model architecture that could integrate physical climate modelling, climate-impact assessment, and techno- and socio-economic analysis within a unified system. The article addresses a long-standing limitation in climate research: although these fields are closely interconnected, they are often analyzed separately using different datasets, modelling assumptions, and computational frameworks.

Current climate assessments typically require complex multi-stage workflows in which Earth system models, impact models, and integrated assessment models are connected through manually coordinated processes. Differences in baselines, assumptions, spatial scales, and model structures can make cross-disciplinary analysis slow and difficult to update. The research team argues that recent advances in artificial intelligence provide an opportunity to redesign this process around a reusable foundation-model architecture capable of learning across multiple climate-relevant domains.

The proposed framework brings together heterogeneous information—including Earth observations, climate and weather simulations, climate–energy–economy model outputs, and regional policy indicators—within a shared latent representation. A Mixture-of-Experts (MoE) architecture would route different inputs to specialized AI modules. This would allow domain-specific expertise to be retained while enabling information to be exchanged across disciplines. Importantly, these expert modules would not need to consist solely of neural networks. Physics-informed components and simplified physical equations could also be incorporated to preserve scientific consistency alongside statistical learning.

This architecture could support policy questions that currently require several models to be run and linked sequentially. For example, assessing how extreme heat affects electricity demand and clean-energy deployment requires information on physical climate extremes, building energy use, demographic vulnerability, and investment behavior. Within the proposed framework, these inputs could be processed together to support applications ranging from climate-risk assessment and infrastructure planning to adaptation strategies and mitigation pathways.

The commentary emphasizes that such an AI system is not intended to replace established Earth system models or integrated assessment models. Rather, it provides a blueprint for connecting and extending them through AI, enabling faster iteration, more consistent assumptions across disciplines, and more flexible analysis of policy-relevant questions. The authors also identify major challenges, including the integration of heterogeneous data, propagation of model uncertainty, explainability, governance, equitable access, and the computational requirements of training large multimodal systems.

The work was published on 28 April 2026 in Nature Climate Change under the title “Artificial intelligence to support cross-disciplinary climate change research.” The article sets out a research agenda for developing the systems over the coming decade and argues that sustained international collaboration will be essential for realizing their potential.

Paper link: https://doi.org/10.1038/s41558-026-02624-x