International roadmap to unite weather and climate prediction through artificial intelligence

The study, coordinated by UV professor Gustau Camps-Valls, identifies three bridges—temporal, causal and structural—that could lead towards a more unified and reliable system for predicting the Earth’s atmosphere, from tomorrow’s storms to changes occurring over decades.

What do weather forecasts for next week and climate projections for the end of the century have in common? Both attempt to describe the evolution of the same Earth system, but they have traditionally been developed by different scientific communities, using different data, assumptions, models and evaluation criteria. This divide has prevented communities from exchanging much experience and tools.

Weather forecasting focuses primarily on the initial state of the atmosphere and on changes over hours, days or weeks. Climate science, by contrast, studies how the Earth system responds over decades or centuries to evolving factors such as greenhouse gas concentrations, aerosols and land-use change. Weather provides scientists with many observable events against which models can be tested; for future climate conditions, there is no equivalent observational record.

The new article argues that artificial intelligence (AI) is beginning to make a leap to narrow this divide. Communities are using very similar techniques, encounter very similar challenges and exchange ideas more actively. And in all these, AI acts as a synergistic catalyst. Techniques such as graph neural networks, transformers, generative models, foundation models, and hybrid physics-AI models are increasingly being used in both fields. They can accelerate simulations, combine large and heterogeneous Earth-observation datasets, increase spatial resolution and generate probabilistic predictions.

The researchers analyzed the evolution of the scientific literature between 1975 and 2025. Their results show that research on weather, climate and AI has grown rapidly and that the thematic overlap between weather and climate studies has increased beyond what would be expected by chance, particularly since the expansion of the deep learning revolution around 2010. Forecasting, downscaling, downscaling, and model emulation are among the methodological areas now shared by both communities.

The paper organizes this convergence around three practical bridges. The first is temporal: subseasonal-to-seasonal prediction covers the intermediate period between conventional weather forecasts and longer-term climate predictions. The second is causal: AI can support the analysis of extreme events under alternative or counterfactual climate conditions, helping researchers examine how climate change alters their probability or intensity. The third is structural: weather and climate models can increasingly share components, representations and hybrid combinations of physical equations and machine learning.

However, the authors stress that predictive accuracy alone is not sufficient. “AI systems must also preserve physical laws, remain stable during long simulations, represent uncertainty and perform reliably when they encounter conditions outside their training data,” says Alberto Carrassi from the University of Bologna. Actually, this is especially important for unprecedented extremes and future climate regimes for which no direct historical equivalent exists.

The work also considers the environmental and social risks associated with AI. Training large models requires energy and computational resources, while access to the necessary infrastructure remains concentrated in a relatively small number of institutions and companies. “AI could make advanced forecasting cheaper to run and more accessible, but it could also widen inequalities if countries and research centers cannot participate in developing the models themselves, as well as cause long-term ifdamage ” states Francisco de Melo Viríssimo, from the London School of Economics and Political Sciences in the UK.

For Camps-Valls, the “central challenge is therefore not simply to design more powerful algorithms, but to establish a new culture of cooperation among meteorologists, climate scientists, physicists, computer scientists, operational forecasting centers, and decision-makers.” Very much in line, “trust will depend on transparency, physical consistency, clear communication of uncertainty and equitable access to data and models,” Kai-Hendrik Cohrs from IPL-UV says.

The study brings together 20 researchers from institutions in Spain, Germany, Italy, the United Kingdom, Greece, the Netherlands and the United States. Alongside Camps-Valls, the UV contribution includes IPL researchers Nathan Mankovich, Esther Rodrigo-Bonet and Kai-Hendrik Cohrs. Camps-Valls designed and structured the study, prepared its first version with the co-authors and supervised the work.

The article originated during the ELLIS workshop “AI for Learning Weather and Climate”, held in València in 2024 within the ELLIS Research Program on Machine Learning for Earth and Climate Sciences. More than 30 specialists met for a week to discuss the scientific and operational barriers separating the two fields.

The research is also connected to major European initiatives in trustworthy AI and Earth-system science. It received support from the European Research Council through the ERC Synergy Grant USMILE, coordinated at the UV by Camps-Valls, as well as from the European projects ELIAS, AI4PEX, MeDiTwin and ThinkingEarth.

Reference: Camps-Valls, G.; Carrassi, A.; de Melo Viríssimo, F. et al. “Bridging the weather and climate divide with artificial intelligence”. Nat Commun 17, 8578 (2026). https://doi.org/10.1038/s41467-026-75787-y