AI Weather Forecasting

AI weather forecasting is the use of machine-learning models to predict the future state of the atmosphere. The leading systems are trained on decades of historical reanalysis data and, from 2023 onwards, have matched or exceeded traditional physics-based numerical weather prediction (NWP) on many measures while running orders of magnitude faster.^[1][2]

History

Weather forecasting has historically relied on NWP: solving the physical equations of the atmosphere on supercomputers. The European Centre for Medium-Range Weather Forecasts (ECMWF) and its Integrated Forecasting System (IFS) have long been regarded as the global benchmark.^[3]

The first wave of AI models appeared in 2023. Pangu-Weather, developed by Huawei Cloud, was published in Nature in July 2023 and was the first AI model reported to exceed the accuracy of traditional numerical methods, producing a 24-hour global forecast in 1.4 seconds on a single GPU — a speed-up of roughly 10,000 times.^[4] Weeks later, Google DeepMind described GraphCast, a graph neural network trained on the ERA5 reanalysis dataset that produced a 10-day global forecast in under a minute on one tensor processing unit and outperformed the operational IFS HRES model on about 90 per cent of evaluated targets, including a nine-day-ahead prediction of Hurricane Lee's landfall.^[2][5]

In December 2024, DeepMind published GenCast, a diffusion model that generates 15-day ensemble forecasts at 0.25° resolution — 50 or more possible weather trajectories rather than a single prediction — in about eight minutes on a TPUv5 device. GenCast outperformed ECMWF's ENS ensemble on 97.2 per cent of 1,320 evaluated targets, including tropical cyclone tracks and extreme conditions, and the model was released openly with its weights.^[6]

ECMWF introduced its own machine-learned system, the Artificial Intelligence Forecasting System (AIFS), operationally on 25 February 2025 — the first fully AI-based system to enter operational service at a major international forecasting centre — followed by an ensemble version, AIFS ENS, with 51 members on 1 July 2025.^[7][8]

Key Concepts and Technology

AI forecasting models learn statistical relationships between successive atmospheric states rather than solving physical equations. They are trained on reanalysis datasets such as ERA5, which reconstruct past weather from observations and historical model runs, and they are initialised from an analysis of current observations.

Architectures vary: GraphCast represents the globe as a multi-scale graph; Pangu-Weather uses a three-dimensional transformer; AIFS uses an encoder-processor-decoder transformer design. Probabilistic systems such as GenCast are diffusion models, adapted to the spherical geometry of the Earth, that sample multiple future scenarios instead of producing a single deterministic forecast.^[2][6]

The main advantages are speed, cost and accessibility: forecasts that once required hours of supercomputer time complete in minutes on a single accelerator, and the models are small enough to be run by national meteorological services, universities and companies. The main limitations are resolution and robustness. AIFS ENS runs at about 31 km resolution against roughly 9 km for the physics-based IFS ensemble, and ECMWF has documented that ML models fine-tuned on a particular analysis can degrade when that analysis is updated unless they are adapted.^[8][9]

Applications and Impact

AI models are used for medium-range global forecasting, tropical cyclone track prediction, flood and disaster early warning, renewable energy (especially wind power), agriculture and logistics. Google's WeatherNext 3, announced in September 2026, added real-time satellite data, hourly updates and higher-resolution precipitation forecasting, and was integrated into consumer products including Google Search, Maps and Gemini.^[10] National meteorological services have deployed their own AI systems, including AIGFS and AIGEFS from NOAA in the United States, AICON in Germany and GEML in Canada. In May 2026, ECMWF stopped running externally developed models — Pangu-Weather, GraphCast, Aurora and FourCastNet — in real time, noting that the field had matured and that many centres now produce their own AI forecasts.^[9]

>See Also

• Time Series Forecasting
• Graph Neural Network
• Digital Twin
• AI in Malaysian Agriculture
• AI Data Centres in Malaysia

References