AI Weather Forecasting

AI weather forecasting is the application of machine learning models to predict atmospheric conditions such as temperature, rainfall, wind, and extreme weather events. Instead of solving the physical equations of the atmosphere at runtime, these models are trained on decades of historical weather data and learn to reproduce atmospheric dynamics directly, producing forecasts of comparable or better skill than traditional numerical weather prediction (NWP) systems while requiring only a fraction of the computing time ^[1].

History

Numerical weather prediction, which simulates the atmosphere by solving mathematical equations on supercomputers, has been the operational standard in meteorology since the 1950s. The first generation of data-driven weather models emerged in the early 2020s, beginning with FourCastNet, a model developed by NVIDIA and collaborators that used adaptive Fourier neural operators to produce global forecasts ^[6].

In July 2023, researchers at Huawei Cloud published Pangu-Weather, a model based on three-dimensional neural networks, in the journal Nature; it was the first AI model to outperform the deterministic forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF) on a range of variables ^[2]. Later that year, Google DeepMind introduced GraphCast, a system built on graph neural networks, which predicted weather conditions up to ten days in advance more accurately than ECMWF's operational High Resolution Forecast (HRES) and produced a ten-day forecast in under a minute on a single TPU device ^[1]. GraphCast was named a runner-up for Science's Breakthrough of the Year in 2023 and received the Royal Academy of Engineering's MacRobert Award in 2024 ^[5].

In December 2024, Google DeepMind published GenCast in Nature, a diffusion-based model that generates probabilistic ensemble forecasts — multiple possible future weather trajectories — and outperformed ECMWF's 51-member ensemble system (ENS) up to 15 days ahead ^[3][4]. In 2024 ECMWF also released its own machine-learning model, AIFS, with open model weights, making AI-based forecasting available to researchers worldwide ^[6]. Subsequent work continued through 2025 and 2026, including Google's WeatherNext 2 and a generative model (FGN) described in a 2026 Nature paper on operational tropical cyclone forecasting with AI ^[5].

Key Concepts

AI weather models are trained on reanalysis datasets such as ERA5, which combine historical observations with model physics to reconstruct the state of the atmosphere over several decades. During training the model learns the statistical relationships between past atmospheric states; at inference time it maps a current state to a future one.

Two broad categories of model exist. Deterministic models such as GraphCast produce a single best estimate of future weather, while probabilistic models such as GenCast sample from a learned probability distribution to generate an ensemble of forecasts, quantifying uncertainty in the same way traditional ensemble systems do but at far lower computational cost ^[3]. GenCast's architecture is based on diffusion models, the same family of generative models used for AI image generation ^[4].

Most leading models operate at a spatial resolution of 0.25 degrees (about 28 km at the equator), covering more than a million grid points over the Earth's surface, with variables predicted at multiple atmospheric levels ^[1]. Because forecasts are produced by a single neural-network forward pass rather than hours of supercomputer simulation, AI models can be run thousands of times more cheaply, enabling rapid updates and large ensembles ^[1][3].

Applications

AI weather models are used for medium-range forecasting, for predicting extreme events such as tropical cyclone tracks, atmospheric rivers associated with flood risk, and the onset of extreme temperatures, and for improving nowcasting and precipitation forecasts ^[1][3]. Google has integrated its models into consumer products including Google Search and Maps, and operates a dedicated flood forecasting system that issues early warnings of riverine flooding in many countries ^[3][8].

Meteorological services and private forecasters increasingly combine AI models with traditional NWP in hybrid systems, and the release of open weights by ECMWF and Google DeepMind has allowed research institutions to run and evaluate the models independently ^[3][6]. The 2026 Nature publication on operational tropical cyclone forecasting marks the growing use of AI models in official early-warning workflows ^[5].

>See Also

• Google DeepMind
• Diffusion Model
• Graph Neural Network
• AI in Malaysian Agriculture
• AI in the Malaysian Public Sector