Technology | Europe
How AI-Powered Climate Modelling Is Transforming Weather Forecasts From Days to Months
AI climate models are now outperforming conventional physics-based models on 10-day forecasts. Here is how this happened and when monthly accurate forecasting might be possible.
AI climate models are now outperforming conventional physics-based models on 10-day forecasts. Here is how this happened and when monthly accurate forecasting might be possible.
- AI climate models are now outperforming conventional physics-based models on 10-day forecasts.
- Nature identified AI-powered climate modelling as one of the seven technologies most worth watching in 2026, and the developments through the year's first quarter vindicate the designation.
- Conventional numerical weather prediction — the approach that has dominated weather forecasting since the 1950s — solves the physical equations that govern atmospheric dynamics numerically.
AI climate models are now outperforming conventional physics-based models on 10-day forecasts.
Nature identified AI-powered climate modelling as one of the seven technologies most worth watching in 2026, and the developments through the year's first quarter vindicate the designation. Google DeepMind's GraphCast, Nvidia's FourCastNet, and several competing AI-based weather prediction systems have collectively demonstrated that machine learning approaches can match or exceed the accuracy of conventional numerical weather prediction on forecasting timescales from one to fourteen days, at a fraction of the computational cost.
Conventional numerical weather prediction — the approach that has dominated weather forecasting since the 1950s — solves the physical equations that govern atmospheric dynamics numerically. This requires enormous computational resources (the most powerful supercomputers in the world are used for operational weather forecasting), trained meteorologists to interpret and correct outputs, and approximately 1-2 percent of global supercomputing capacity. The physical equations encode genuine physical laws, which gives the approach interpretability and a principled foundation.
AI-based models learn from historical weather data without explicit physical equations — they identify statistical patterns in how weather evolves over time and use those patterns to predict future states. The surprising finding is that these statistical patterns are rich enough to produce forecast accuracy comparable to physics-based models for periods up to two weeks. For some specific variables — sea surface temperature, precipitation intensity — AI models now outperform conventional models.
The potential breakthrough that would most change practical life: subseasonal-to-seasonal (S2S) forecasting — accurate predictions 2-8 weeks ahead. Current conventional models lose significant accuracy beyond 10 days as chaotic atmospheric dynamics amplify small errors. Whether AI models can extend meaningful forecast accuracy into the S2S range by learning longer-range statistical patterns from historical data is the frontier question. Preliminary results from AI S2S forecasting are promising but not yet at operational accuracy levels.
For agriculture, energy systems, disaster preparedness, and supply chain management: a meaningful improvement in 3-6 week forecast accuracy would be economically transformative. Farmers could make planting and irrigation decisions with better information. Energy grid managers could anticipate supply-demand mismatches from weather-dependent renewable generation. Emergency managers could prepare for severe weather events two weeks further in advance.