Physics-Informed Deep Learning for Radar Precipitation Nowcasting

A team of researchers has introduced a deep learning framework informed by physics to assess altitude-specific motion fields using volumetric radar reflectivity data for precipitation nowcasting. This model employs a fully differentiable semi-Lagrangian extrapolation operator, allowing it to handle 3D data as separate horizontal slice sequences. This approach facilitates efficient inference of horizontal motion at various altitude levels. When tested on a multi-year radar dataset from Central Europe, the motion fields produced demonstrated significant vertical coherence and a high correlation across altitudes, resulting in enhanced forecasts based on extrapolation.