Artificial-intelligence models are beginning to supplement and in some cases outperform traditional numerical weather prediction (NWP) models, which divide the atmosphere into a three-dimensional grid and crunch through equations on supercomputers. AI weather models work differently: rather than simulating atmospheric physics equation by equation, they compare patterns in current weather data with historical records on which they have been trained.
In 2022 Nvidia published the first results from FourCastNet, an AI weather program trained on decades of weather data. The company claimed it accurately predicted hurricanes and rainfall a week in advance with just two seconds of computing time—thousands of times less than what an NWP model needs. Tech firms and weather agencies began racing to build their own.
Relying solely on AI came with problems. Unconstrained by the laws of physics, forecasts could become unrealistic. Fed only past data, the models often struggled to predict rare or extreme events. The focus therefore shifted to combining the best of both approaches. The European Centre for Medium-Range Weather Forecasts (ECMWF), widely considered the world's best forecasting institute, developed an AI model to work alongside its existing NWP software. Google's NeuralGCM still grinds through calculations to represent big atmospheric processes but uses AI to fill in the details.
Pedram Hassanzadeh, an AI and extreme-weather researcher at the University of Chicago, argues that AI models could lead to a "democratisation of weather forecasting". NWP forecasts demand supercomputers and many weather-observing stations; poor countries often lack both. The Human-Centered Weather Forecasts Initiative (HCF), which Dr Hassanzadeh helps lead, encourages governments to use AI to overcome these barriers. Once trained, AI weather models can run on a high-end laptop and can be tweaked to focus on the best-quality data available locally.
In 2025 the Indian government sent about 38m farmers forecasts generated by AI models rather than NWP ones—using an ECMWF model and a version of Google's NeuralGCM. The models predicted when the monsoon rains would arrive 30 days ahead in some regions, and forecast a 20-day stall in the middle of the season that did not appear in NWP forecasts. Almost half the farmers who received the messages said the information influenced their planting decisions.
In October 2025 the ECMWF announced it would make both its most up-to-date forecasts and its assimilated data freely available. Like NWP models, AI models still depend on "data assimilation"—the process of converting messy observations from satellites and weather stations into a well-ordered snapshot of the atmosphere. That work remains concentrated among rich-country agencies.
The Gates Foundation and the United Arab Emirates fund the HCF's work to extend AI weather forecasting to east and west Africa. The HCF is teaching meteorological and agricultural officials from Bangladesh, Chile, Ethiopia, Kenya and Nigeria how to use AI weather models for their own particular needs.
It destroys one's nerves to be amiable every day to the same human being.