Artificial intelligence (AI) has already demonstrated its proficiency at difficult scientific tasks like predicting how proteins will fold and identifying new astronomical objects in masses of observational data [
1]. Now, recent results suggest that AI also excels at weather forecasting. For global predictions, GraphCast, an AI system developed by Google subsidiary DeepMind (London, UK), outperforms the state-of-the-art model from the European Centre for Medium-Range Weather Forecasts (ECMWF), providing more accurate projections of variables such as temperature and humidity
of the time [
2,
3]. Other AI systems, including Pangu-Weather from the Chinese tech company Huawei (Shenzhen, China) [
4], can also match or beat traditional global forecasting models.
The speed of improvement in AI’s weather prediction ability has surprised meteorologists. "What has happened with AI is astonishing," said Dale Durran, professor of atmospheric sciences at the University of Washington, Seattle, WA, USA. As a result, it is not a question of when AI will provide forecasts for people to use. "We are already there," said Aaron Hill, an assistant professor of meteorology at the University of Oklahoma, Norman, OK, USA. Google’s nowcast feature for mobile devices, for instance, offers AI precipitation predictions for the next 12 hours [
5]. And ECMWF has incorporated AI into its leading model [
6]. What is uncertain, however, is how big a role AI will play in weather forecasting. Will it be one of several tools, or will it replace conventional models-or even put human forecasters out of work [
7]? These questions are important because accurate weather forecasting is not just helpful for shaping day-to-day human activity; it is vital for preparing responses to mitigate the damage caused by extreme weather (
Fig. 1).
Traditional weather forecasts such as those generated by the US National Weather Service (NWS) and the ECMWF rely on numerical models that use large amounts of data on current conditions to project how atmospheric conditions will evolve [
8,
9]. Researchers have developed many of these models that deliver results for different time and spatial scales and for a variety of applications. The atmospheric model at the heart of the ECMWF’s Integrated Forecasting System (IFS), for instance, divides the global atmosphere into a grid of several billion boxes that are
on a side and then predicts changes in variables such as wind speed and temperature in each cell [
10,
11]. IFS delivers a single,"high-res" ten-day forecast as well as longer-term predictions. Other models give even greater geographic precision. The High-Resolution Rapid Refresh model from the US National Oceanic and Atmospheric Administration (NOAA) offers two-day forecasts with a resolution of
[
12], but it only covers the United States [
12].
The accuracy of forecasts made with numerical models has increased dramatically in the last few decades. Today’s six-day forecasts, for example, are correct about as often as three-day forecasts were in the early 1990s [
2]. Two reasons for this improvement are better weather models and faster, more powerful computers, said Durran.
But these advances have come at a cost. Today’s weather models are so complex that they require supercomputers to run them-which makes standard forecasting an expensive task. In 2020 the ECMWF paid more than 90 million USD for a BullSequana XH2000 supercomputer from the French company Atos (Bezons, France), and the UK Met Office doled out 1.5 billion USD for a new supercomputer built by Microsoft (Redmond, WA, USA) [
13,
14]. The models also take a relatively long time to produce their results. The IFS, for instance, updates its forecasts only four times a day [
15]. And despite their proven abilities, numerical models still struggle to predict certain kinds of weather events, such as thunderstorms [
16].
Interest in applying AI to weather forecasting is not new. For decades, academic and government scientists have been investigating the capabilities of machine learning and other types of AI, said Auroop R. Ganguly, professor of civil and environmental engineering at Northeastern University in Boston, MA, USA. However, only recently have tech powerhouses like Google, Microsoft, and Huawei become involved in the research [
7]. "The industry-scale use is very new," Ganguly said, and it is the main driver of the rapid improvements in AI’s forecasting accuracy [
7].
Unlike standard models, AI systems such as GraphCast and Pangu-Weather do not attempt to numerically simulate the atmosphere. Instead, researchers train the systems by feeding them large amounts of historical weather data-GraphCast learned from 39 years’ worth of data, covering 1979 to 2017 [
3]. The AI systems pore over this data to identify patterns and then make predictions by comparing current conditions to past weather [
7]. The resulting forecasts are faster and cheaper to generate than those made by traditional models. GraphCast can produce a ten-day forecast comparable to the IFS high-res forecast in less than a minute [
3]. And once an AI system has completed its training period, it can be run on a desktop computer, dramatically reducing the costs of making forecasts [
2].
Meteorologists describe a model that produces accurate predictions as skillful, and "there is a lot of skill in these AI models," said Durran. To evaluate GraphCast, which makes its predictions from the current conditions and the conditions
previously, the researchers who created the model gauged its ability to forecast variables like temperature, wind direction, and humidity at 13 different levels in the atmosphere over a ten-day period. Overall, predictions from GraphCast were more accurate than those from the IFS high-res forecasts
of the time [
3]. In addition, GraphCast beat the ECMWF model at projecting the tracks of tropical cyclones, identifying atmospheric rivers that carry large amounts of moisture, and predicting episodes of extreme heat and cold [
3]. By performing a similar analysis, researchers showed that Pangu-Weather was also superior to the IFS on the seven-day outlook [
17].
AI systems still fall short of numerical models in some ways. GraphCast and Pangu-Weather only reach a resolution of about
, versus
for the IFS. And the AI models sometimes miss key weather events. For instance, a 2024 study by researchers at the University of Reading in Reading, UK, assessed whether four AI models, including GraphCast and Pangu-Weather, could forecast (retrospectively) Storm Ciarán, a hurricane-like storm that tore through Europe in late 2023, killing 16 people [
18]. The models did capture some aspects of Ciarán’s structure, but all underestimated the speed of the winds, one of the most destructive features of the storm, which reached
[
18].
Although the AI results are impressive, the models "are in their infancy," said Hill, and scientists have just begun testing them to identify their strengths and weaknesses. Researchers need to know more about their performance, and "they will not replace the traditional models soon," he said. But Durran disagreed. "It is extremely likely this is going to happen," he said. The AI models already do roughly as well as global numerical models, and developers should be able to correct the remaining deficiencies, he said.
Experts have raised several concerns about relying on AI to forecast weather. One worry is AI systems’ tendency to hallucinate, or invent results, a common problem with generative AI chatbots like ChatGPT [
19]. If AI models predict non-existent weather events, the results could be disastrous. Researchers do not have enough information to say whether the models are hallucinating, said Hill. Most current models sometimes produce unrealistic results, but it is possible to create models that do not, Durran said. He said a model he and his colleagues developed can simulate weather patterns for 100 years into the future-assuming current climatic conditions hold-without hallucinating [
20]. All the weather conditions the model predicts are "similar to what we see in the range of weather patterns today," he said. Another concern is whether models trained with historical weather data will be able to make accurate predictions as the climate changes [
21]. However, feeding the models recent data may overcome this problem, according to some researchers [
3].
Ganguly said that a key issue that will help determine whether AI models gain wider acceptance is how valuable their predictions are for users of weather information. Operators of hydroelectric dams, for example, need accurate forecasts to prepare for flooding or other weather-related threats, he said. If an AI system can provide better information and improve operators’ ability to make decisions, they will be more likely to use it.
Some experts have suggested that AI could even replace human forecasters [
7], an outcome that other experts consider unlikely. To make their forecasts, humans typically weigh results from multiple models and then incorporate their knowledge of local weather patterns. AI cannot match the combination of judgment and knowledge, Hill said. "That experience cannot be replicated," he said. At least for now.
PDF
(471KB)
