Scientists have developed a machine learning model that can outperform official agencies at predicting tropical cyclone tracks, and do it faster and cheaper than traditional physics-based systems.
Aurora, a foundation model developed by researchers from Microsoft, the University of Pennsylvania (UPenn), and several other institutions, is designed to improve the speed and accuracy of Earth system forecasts, from air quality and ocean waves to tropical cyclone tracks and high-resolution weather.
Paris Perdikaris, co-lead author and associate professor of mechanical engineering and applied mechanics at UPenn, describes Aurora as a large neural network. Like ChatGPT does for text, Aurora learns from past geophysical data to predict complex physical processes, without relying explicitly on traditional physics equations.
“Traditional models are designed on first physical principles, like conservation of mass, momentum and energy,” he said. “Aurora, on the other hand, is not directly using those physical principles, but instead it relies on observations and data.
“Aurora learns from a very diverse set of geophysical data, including forecasts, observations, and what we call analysis and re-analysis data, which is basically a reconstruction of historical weather patterns we have access to.”
As the researchers acknowledge in the Nature paper published Wednesday, Aurora’s rapid progress depends heavily on the groundwork laid by traditional methods. “Such an accelerated timeline is only possible because of the wealth of data that is available as a result of decades of research into traditional numerical approaches.”
Aurora was pretrained on more than one million hours of diverse geophysical data and fine-tuned over four to eight weeks by small engineering teams, “compared with a typical development period of several years for dynamical baseline models,” the paper noted.
Aurora, trained only on historical data, was able to correctly forecast all hurricanes in 2023 more accurately than operational forecasting centers
The model uses a combination of Perceiver-based encoders, a 3D Swin Transformer backbone, and recursive forecasting techniques, relying on multi-dimensional vector embeddings similar to those used in large language models.
“Aurora, trained only on historical data, was able to correctly forecast all hurricanes in 2023 more accurately than operational forecasting centers,” Perdikaris touted.
The authors reported that Aurora outperformed seven operational forecasting centers on five-day tropical cyclone track predictions for all global cyclones in 2022–2023. It also surpassed state-of-the-art numerical models on 92 percent of targets in 10-day global weather forecasts at 0.1° resolution.
As a foundation model, they suggest Aurora could be fine-tuned for a wide range of Earth system prediction tasks beyond weather, including air quality, ocean dynamics, and environmental extremes.
Aardvark beats groundhogs and supercomputers in weather forecasting
Microsoft: So what if it costs 4X as much to run Windows Server in AWS, Alibaba, and Google?
Google DeepMind touts AI model for ‘better’ global weather forecasting
US govt hiding top hurricane forecast model sparks outrage after deadly Helene
“The potential implications of Aurora for the field of Earth system prediction are profound,” the authors wrote. “Although in this paper we showcase the application of Aurora to four domains, it could be fine-tuned for any desired Earth system prediction task, potentially producing forecasts that outperform the current operational systems at a fraction of the cost.”
Potential applications range from modeling ocean currents, short- and long-term weather patterns, and vegetation cycles, to forecasting wildfires, floods, crop yields, pollination behavior, renewable energy output, and shifts in sea ice coverage.
“With the ability to fine-tune Aurora to diverse application domains at only modest computational cost, Aurora represents notable progress in making actionable predictions accessible to anyone,” they added.
It’s not the only AI model shaking up meteorology. In March, Aardvark, a novel machine learning-based weather prediction system, showed promise in outperforming traditional supercomputer-based models. The system can be trained and run on a desktop equipped with NVIDIA GPUs, generating 10-day forecasts in minutes, and at a fraction of the computational cost of current numerical weather models. ®