A new weather forecasting computer model created by Google, powered by artificial intelligence, consistently outperforms and is several times faster than government models that have been around for decades and invested hundreds of millions of dollars, a study published on Tuesday showed.
Google’s AI weather forecasting model is surprisingly accurate, study finds
The study, published in the journal Science, showed that the AI model is more accurate for daily weather forecasts and extreme events, such as hurricanes and extreme heat and cold.
Its excellent performance and promising results from other AI models like it may signal the beginning of a new era for weather forecasting, although experts say this does not mean AI is ready to replace all traditional forecasting methods.
Google DeepMind’s AI model, called “GraphCast,” has been trained on nearly 40 years of historical data and can deliver 10-day forecasts at six-hour intervals for locations spread around the world in less than a minute on a computer the size of a computer screen. small box. Whereas the traditional model would take an hour or more on a school bus-sized supercomputer to accomplish the same feat. GraphCast was more accurate than the European model for more than 90 percent of the weather variables evaluated.
The study’s findings are similar to those reported in an academic article published in August on the online database arXiv.
“Being able to compete with, if not outperform, the world’s best forecasting system is amazing,” Aaron Hill, lead developer of the machine-learning forecasting system at Colorado State University, said in an email. “You can safely add GraphCast to the growing list of AI-based weather forecasting models that should see continued evaluation for their applications in industry, research and operational forecasting.”
AI-based weather models have attracted increasing interest from government meteorological agencies because of their speed, efficiency, and potential cost savings.
Traditional weather models, such as the “European” ones operated by the European Center for Medium-Range Weather Forecasts (ECMWF) in Reading, UK The American National Oceanic and Atmospheric Administration makes predictions based on complex mathematical equations. Such models support life-saving forecasts and warnings around the world, but they are expensive to run because they require huge amounts of computing power.
AI models use a different approach. They are trained first to recognize patterns in vast amounts of historical weather data, then generate forecasts by assimilating current conditions and applying what they have learned from historical patterns. This process is less computationally intensive and can be completed in minutes or even seconds on much smaller computers.
The ability to learn from growing archives of past weather data is a key advantage of AI models. “This has the potential to improve prediction accuracy by capturing patterns and metrics in the data that cannot be easily represented in explicit equations,” the authors who developed the model wrote in the study.
The performance of GraphCast was evaluated against the European model not only for individual weather variables such as temperature, wind and pressure, but also in forecasting extreme events including tropical cyclones, atmospheric rivers, heat waves and cold waves.
Researchers have expressed concerns about AI’s ability to accurately predict severe weather, in part because there have been relatively few such events to learn from in the past. However, GraphCast reduced hurricane forecast errors by about 10 to 15 miles in a two- to four-day time period, improved forecasts of water vapor associated with atmospheric rivers by 10 to 25 percent, and provided more accurate forecasts of extreme heat and cold by five to five. 10 days before the scheduled date.
“Conventional wisdom would say that using (AI) might not be good for rare and unusual things. It seems like it was good for this,” Peter Battaglia, director of research at Google DeepMind and one of the study’s co-authors, said in an interview. “We think this also points to the fact that the model is capturing something more fundamental about how weather actually evolves over time rather than just looking for more superficial patterns in the data.”
Hill cautions that while the study “reinforces the idea that for most events, skillful predictions can be made,” the results do not remove questions about the effectiveness of artificial intelligence in predicting extreme events. “The study describes some fairly broad and comprehensive statistics for extreme weather forecast skill, which indicate how well a model performs in many events, but does not necessarily provide details about how it performs in any one extreme event,” he said.
Other challenges remain before AI models such as GraphCast can be used reliably in operational forecasting. For example, due to limitations in training data and engineering constraints, global AI models are not yet able to generate predictions for as large a number of parameters or as detailed as those found in traditional models. This makes AI models less useful in forecasting smaller-scale phenomena, such as thunderstorms and flash floods, or larger weather systems that can produce large variations in precipitation amounts over small distances.
Meteorologists must also learn to trust AI models, whose inner workings are less transparent than traditional models.
“The main role of forecasters is to interpret information and communicate it to partners, a task made more difficult by the lack of tools to determine why an AI model predicts what it does,” said Jacob Radford, a data visualization researcher at the Collaborative Institute for Forecasting Science. Atmospheric research at Colorado State University said in an email. “These models are still in their infancy and confidence still needs to develop in both the research community and forecasters before practical use can be considered.”
Most experts, including the study’s authors, agree that traditional models are not about to be replaced by AI models, which still rely on traditional models to provide training data and generate current conditions that they use as a starting point for predictions.
“Our approach should not be considered an alternative to traditional weather forecasting methods, which have been developed over decades, have been rigorously tested in many real-world contexts, and offer many advantages that we have not yet explored,” the authors wrote. “Instead, our work should be interpreted as evidence that (AI weather forecasting) is capable of meeting the challenges of real-world forecasting problems, and has the potential to complement and improve upon the best existing methods.”
Recent advances in weather forecasting AI
Major technology companies, including China-based Google, Microsoft, Nvidia, and Huawei, have made rapid progress in weather modeling using artificial intelligence in the past two years. All four companies have published academic articles claiming that their global AI models perform at least as well as the European model. These claims were recently confirmed by scientists at ECMWF.
In September, artificial intelligence models developed by Google, Nvidia, and Huawei successfully predicted Typhoon Lee’s path a week before it occurred. The hurricane quickly intensified into a Category 5 hurricane in the Atlantic Ocean east of the Caribbean, then weakened before eventually making landfall in Nova Scotia with strength equivalent to a tropical storm.
ECMWF began publishing forecasts from all three models on its website just one day after Lee first developed into a tropical storm. NOAA’s Cooperative Institute for Atmospheric Research at Colorado State University will launch a similar website by early December, according to Radford.
Meanwhile, the UK Met Office recently announced a collaboration with researchers at Britain’s Alan Turing Institute to develop artificial intelligence forecast models to “improve the prediction of some extreme weather events, such as exceptional rainfall or impact thunderstorms, with greater accuracy.” The Met Office said in a press release.
Earlier this month, Google announced another AI model that can provide more local forecasts of precipitation, temperature and other parameters for up to 24 hours using direct observations from weather sensors as a starting point.
Besides modelling, AI is also being used to enhance communication and interpretation of weather forecasts. The National Oceanic and Atmospheric Administration (NOAA) announced last month that it is using artificial intelligence to automate the translation of weather forecasts into Spanish and Chinese, with additional languages to follow, while private weather company Tomorrow.io has developed an AI assistant called “Gale” to help business clients. On the interpretation of the weather. Expectations for specific use cases.
