Weather forecasting relies on complex computer simulations that can take hours to run. Two new AI approaches may help meteorologists work faster and improve some predictions, including warnings about extreme rain. Researchers say the systems could complement conventional forecasts, though their performance in everyday use is still an open question.
Forecasting many weather variables at once
Traditional forecasting systems analyze variables such as temperature, precipitation, pressure, wind, humidity, and cloudiness. Huawei’s Pangu-Weather takes a different approach: its deep neural network processes all of them at the same time, producing a forecast in seconds.
The model was trained on 39 years of reanalysis data, which combines historical weather observations with modern models. In tests against the operational integrated forecasting system of the European Centre for Medium-Range Weather Forecasts (ECMWF), Pangu-Weather produced similar accuracy while working much faster.
Researchers also found that it could accurately track a tropical cyclone’s path, even though its training data did not include tropical cyclones. Oliver Fuhrer of MeteoSwiss, who was not involved in the research, said this suggests machine-learning models can learn physical processes and apply them to situations they have not encountered in training.
That result points to a possible role for AI alongside established forecasting methods. Pangu-Weather does not need to replace conventional systems to be useful: faster analysis could give forecasters another way to assess what may happen and help authorities prepare. Huawei researcher Lingxi Xie says the models could be used together with conventional methods.
More time to prepare for heavy rain
A separate approach targets a particularly difficult problem: predicting extreme rainfall early enough for people to respond. NowcastNet, a physics-based generative AI model, can predict extreme rain up to three hours in advance. The article says that is a longer lead time than existing conventional methods.
In evaluations by sixty-two Chinese meteorologists, NowcastNet ranked as the best rain prediction method in around 70% of cases against similar systems. For comparison, DeepMind’s DGMR predicts the likelihood of all rain in the next 90 minutes, while NowcastNet focuses on the harder task of identifying extreme rain.
The researchers trained NowcastNet on data from weather radars and other technologies, including sensors and satellites. They also built in principles of atmospheric physics, such as gravity. Radar data gives the model snapshots of weather patterns, which it uses to generate a likely next scenario.
That combination may help the system handle rare events. The article describes radar-only models as having a partial snapshot of the atmosphere, which can make extreme rainfall harder to predict. By incorporating physical principles, NowcastNet aims to form a fuller picture of how rain may behave. Better short-term warnings could give people more time to prepare for a type of event that can cause death and destruction.
Limits remain, especially for extremes
The reported results do not mean AI can predict every dangerous feature of a storm. Pangu-Weather can help track where tropical cyclones are heading, but it cannot forecast how intense they will be. Xie says AI tends to underestimate extreme weather, a limitation that matters when a forecast is being used to prepare for severe conditions.
There is also a difference between the data used to train Pangu-Weather and the observational data used by conventional weather prediction systems. Pangu-Weather relies on reanalysis data, and Xie says the team hopes to train it on observational data in the future. Whether that change would improve its performance is not established in the source.
AI weather forecasting is still at an early stage. Pangu-Weather, Nvidia’s FourcastNet, and Google-DeepMind’s GraphCast have prompted meteorologists to reconsider how machine learning might fit into forecasting. But researchers still need to see how these systems perform in practice.
A changing climate adds another uncertainty. Peter Dueben of ECMWF raised the question of how a model such as Pangu-Weather would respond to conditions unlike those it has seen, including a future in which all the ice in the Arctic disappears. Faster forecasts and better rain predictions could help, but their value will depend on how reliably the systems handle real and changing weather.