Recently, scientists have developed a machine learning model named "Aurora," which outperforms official agencies in predicting tropical cyclone trajectories while being faster and more cost-effective. Aurora is a foundational model co-developed by researchers from Microsoft, the University of Pennsylvania, and other institutions. It aims to enhance the speed and accuracy of Earth system predictions, covering areas such as air quality, ocean waves, tropical cyclone trajectories, and high-resolution weather.
Image Source Note: Image generated by AI, licensed by MidJourney.
Paris Perdikaris, an associate professor of mechanical engineering and applied mechanics at the University of Pennsylvania and a co-author of Aurora, stated that Aurora resembles a large neural network capable of learning from past geophysical data to predict complex physical processes without relying on traditional physical equations. Traditional models are based on fundamental physical principles like mass, momentum, and energy conservation, whereas Aurora learns through observation and data.
Aurora underwent pre-training with over a million hours of diverse geophysical data and was fine-tuned within four to eight weeks with the assistance of small engineering teams. This process is significantly faster and more efficient compared to traditional dynamical models, which typically require years of development cycles.
According to the research team's report, Aurora accurately predicted all hurricanes in 2023 and outperformed current meteorological forecasting centers. Additionally, the model surpassed seven operational forecast centers in five-day tropical cyclone trajectory predictions from 2022 to 2023 and outperformed 92% of targets in ten-day global weather forecasts.
As a foundational model, Aurora has broad application potential beyond weather prediction, including air quality, ocean dynamics, environmental extreme events, and more. Researchers noted that the emergence of Aurora could have profound impacts on the field of Earth system predictions, providing more accurate forecasts at lower costs.
Meanwhile, another machine learning weather prediction system called "Aardvark" is also gaining prominence. Aardvark demonstrates the potential to outperform traditional supercomputer models, capable of training and running on desktops equipped with NVIDIA GPUs, generating ten-day weather forecasts at lower computational costs.
Key Points:
🌪️ The Aurora model outperforms official agencies in predicting tropical cyclone trajectories, with faster speed and lower cost.
📊 The model underwent over a million hours of data pre-training, with a fine-tuning period of only four to eight weeks.
🌍 The potential applications of Aurora include air quality, ocean dynamics, and environmental extreme events among other fields.