Is AI the future of weather and climate modelling?

The short answer: we don¡¯t know. AI is developing fast, so even experts cannot reasonably predict where it is heading. The longer and more nuanced answer would be: not likely. While there are good reasons to believe that AI will play a rapidly increasing role, we argue here that some aspects of climate and weather modelling will never be fully handed over to AI. Let us explain why.

Weather prediction: the AI success story

AI models outperform traditional forecasting systems for specific weather predictions on lead times from hours to seasons. Why? Because weather offers exactly what AI thrives on: data, and lots of it.

Thanks to over 80 years of global observations and reanalysis datasets, nota-bene primarily obtained and paid for by the public, scientists can feed AI systems with millions of examples of how the atmosphere behaves. These models learn patterns and relationships from this vast amount of data that are inconceivable for humans. In just a few years, AI has gone from being something of experimental curiosity to a genuine competitor for conventional weather models. AI forecasting models offer more accurate, faster and more cost-effective forecasts, which is particularly beneficial for regions with limited access to supercomputing resources and helps to democratise weather forecasting.

We thus expect the application of AI-based methods to continue to thrive in this field. 

Climate, however, poses an entirely different challenge. We only have one observed climate record. That makes it very difficult for AI to train on the diverse datasets it thrives on. Even more daunting is the fact that future climate conditions will fall outside the range of past observations.

Climate modelling: a tougher nut to crack

Slow-changing systems such as deep ocean currents, ecosystem shifts and melting ice sheets are notoriously difficult to model, even with the most sophisticated physics-based tools. Feedback loops¡ªlike the melting of ice shelves or shifts in ocean circulation¡ªare also extremely difficult for AI to anticipate. These tipping elements can trigger massive, rapid changes after decades of seeming stability.

AI models, often trained to minimise error under ¡°normal¡± conditions, have been shown to struggle when rare, extreme, or unseen situations occur¡ªthe very moments when reliable information is needed most.

The future: collaboration, not competition

The most realistic scenario we see is not one where AI replaces traditional numerical models, but one where it complements and enhances them. AI has many more applications than just emulating a numerical model. It can help speed up numerical simulations, analyse massive datasets and extract information, not to mention improve specific model components such as cloud formation and representations of turbulent systems like thunderstorms.

An exciting frontier is the development of digital twins of the Earth¡ªhigh-resolution virtual replicas that combine physics-based models with AI to simulate and predict Earth system behaviour. Initiatives like NVIDIA¡¯s Earth-2 and the EU¡¯s Destination Earth aim to provide real-time insights into extreme weather, climate risks and adaptation options. These efforts exemplify how AI and traditional modelling can merge to support decision-making and improve societal resilience.

Although progress has been made in explainable AI (XAI), the gold standard for scientific understanding remains grounded in physics, theory and observation. Unlike AI models, physics-based approaches provide interpretability, traceability and insight into why a system behaves the way it does. If AI technologies are to serve the public good, we need strong education, public engagement and ethical frameworks.

So, will AI define the future of climate and weather modelling? AI is an extraordinary tool, but not a magic bullet. We believe AI will help shape the future, but it won¡¯t define it alone. Understanding, predicting and managing our planet¡¯s complex systems will still rely on physical foundations and the creativity and collaboration of human scientists.