Neural Networks Could Shed Light on High-Level Clouds
Scientists at Tomsk State University have developed a method for reconstructing the properties of high-level clouds by combining atmospheric optics with machine learning.
The new method could improve weather forecasting and help provide earlier warnings of dangerous cyclones.
Researchers at the university’s High-Energy Physics Data Analysis Laboratory combined lidar, satellite and atmospheric sounding data with machine-learning tools. They found that directly training neural networks on raw meteorological data was ineffective: the atmosphere is a dynamic system, and the algorithm needs to account for interactions among temperature, humidity, pressure and wind at different altitudes. The researchers therefore converted the 2009–2023 ERA5 reanalysis archive into compact digital codes.
A second-year master’s student at Tomsk State University proposed the best neural-network architecture in a machine-learning competition. He has since joined the research team and is refining the model for deployment.
