Reconstructing 4D Wind Fields from Radar Observations using Machine Learning

One of the main factors characterizing the dynamics in the atmosphere is its vertical density stratification. Gravity waves propagation upwards and breaking in the middle atmosphere play an essential role in large-scale energy transport, planetary-scale circulation and the generation of stratified turbulence, manifesting in phenomena such as the cold summer mesopause in the mesosphere. Direct observation or numerical simulation of these processes with high resolution proves difficult however due to the remoteness of the region combined with horizontal scales of 10-100km and vertical scales of 10-100m that have to be resolved for a detailed analysis of the underlying stratified turbulence.

To tackle these limitations and further our knowledge on turbulence activity in the middle atmosphere, we combine the physics-informed machine learning method HYPER (Hydrodynamic Point‐wise Environment Reconstructor) with state-of-the-art radar observations from MAARSY (Middle Atmosphere Alomar Radar System) and SIMONe (Spread-spectrum Interferometric Multistatic Meteor Radar Observing Network). The method allows reconstruction of complete 4D wind fields (spatial+temporal) based on line-of-sight measurements while adhering to Navier-Stokes-based physics constraints and has been successfully deployed previously to extract winds on 10km-scales from inputs of SIMONe. 

In our work we extend the procedure to combine the input of MAARSY and SIMONe and predict complete 4D wind fields at unprecedented horizontal and vertical resolution. Using DNS of stratified turbulence with virtual radars as a validation case, we show that our improved method is able to produce accurate results in the entire prediction domain beyond the provided measurement points, while respecting the given physics constraints. Building on this, we aim to provide a first machine learning supported analysis of stratified turbulence in the mesopause region based on radar observations.