How AI Is Changing Climate and Weather Forecasting

Summary

Carbon Brief interviewed 10 scientists on how AI is being used across climate and weather research
AI models like GraphCast can match or beat traditional physics-based forecasts using far less computing power
GraphCast outperformed Europe's leading forecasting system on about 90 percent of tested metrics
Traditional models still have an edge in predicting record-breaking extreme weather events
Experts warn of risks including energy use, unverified outputs, and unequal access to computing resources

Ten climate scientists surveyed by Carbon Brief say artificial intelligence is reshaping weather and climate research, offering faster, cheaper forecasts but raising concerns about reliability and energy use.

Artificial intelligence is no longer confined to chatbots and image generators. It is now quietly reshaping how scientists forecast the weather and study the climate, according to a survey of 10 experts published by Carbon Brief on 22 October. The scientists, who work across fields from weather prediction to energy-grid analysis, described AI tools that are faster, cheaper and in some cases more accurate than the physics-based models that have dominated forecasting for decades.

Forecasts in minutes, not hours

Traditional weather prediction relies on supercomputers solving complex mathematical equations that describe how the atmosphere behaves. AI models take a different route. Dr Doug McNeall of the UK Met Office explained that these systems learn patterns from large volumes of historical weather data, then generate forecasts by feeding their own predictions back into themselves repeatedly, producing outlooks up to roughly two weeks ahead.

The clearest example is GraphCast, a neural-network model built by Google DeepMind researcher Rémi Lam and trained on four decades of atmospheric data. In its original study, GraphCast outperformed the European Centre for Medium-Range Weather Forecasts’ leading operational model on around 90 percent of more than 1,380 verification targets. It can produce a 10-day global forecast in under a minute on a single processor, compared with hours of supercomputer time for conventional methods.

Deep learning offers a different approach, using data instead of physical equations to create a weather forecast system.

That quote comes from Rémi Lam himself, speaking to Climate Foresight, the magazine of the Fondazione CMCC. A 2025 study from the University of Washington went further, finding that refining the atmospheric data fed into GraphCast could extend useful forecast accuracy to as far as 33 days ahead, roughly a month. The researchers were careful to note that more testing is needed before such long-range AI forecasts can be trusted operationally.

Where AI still falls short

Despite these gains, AI is not yet outperforming traditional methods across the board. Research published in Science Advances found that physics-based models still do better at forecasting record-breaking extreme weather, such as unprecedented heatwaves, cold snaps or windstorms, events that by definition fall outside the historical data AI systems are trained on.

The experts interviewed by Carbon Brief also flagged broader risks. These include scientists placing too much trust in AI outputs without independent verification, the significant energy consumption required to train and run large AI models, and the danger that well-resourced institutions such as Google DeepMind or ECMWF could pull further ahead of smaller research centres with less access to computing power, potentially widening inequalities within climate science itself.

Even so, the direction of travel is clear. Rémi Lam was named one of Nature’s Top 10 scientists of 2024 for his work linking AI to weather and climate forecasting, and similar models, including Microsoft’s Aurora and the Pangu-Weather system, are now part of a growing toolkit that forecasters are testing alongside, rather than instead of, their traditional methods.

This article is free to read. It always will be — no paywall, no account, no tracking.