Augmented Four‐Dimensional Mesosphere and Lower Thermosphere Wind Field Reconstruction via the Physics‐Informed Machine Learning Approach HYPER

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Urco, Juan M., Feraco, Fabio, Chau, Jorge L., Marino, Raffaele
• Format: Journal Article
• Language: English
• Published: American Geophysical Union/Wiley 01.09.2024
• Subjects: Physics
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000162

The mesosphere and lower thermosphere (MLT) is a fluid framework whose multiscale dynamics is determined by a superposition of non‐linear processes and by the interplay of gravity waves and turbulent motions. A thorough comprehension of this atmospheric region requires substantial observational infrastructure, needed to resolve and disentangle its complex dynamics. State‐of‐the‐art observational methods struggle to accurately capture mesoscale dynamics due to the inherent difficulty to perform observations at MLT altitudes. A majority of the observational methods rely on assumptions such as homogeneity, smoothness of the prognostic fields, or zero vertical wind velocities, which may not hold in the upper atmosphere at the mesoscales. In this study, we introduce a novel machine learning‐based approach HYPER (HYdrodynamic Point‐wise Environment Reconstructor), designed to characterize MLT dynamics. HYPER utilizes a physics‐informed neural network to project sparse Doppler meteor detections into four‐dimensional time‐series arrays containing the Cartesian components of the velocity field. This method combines meteor radar observations with the physics prescribed by the Navier‐Stokes equations. The validation of HYPER was conducted through a series of benchmarks on numerical data and the application of our algorithm on actual meteor radar observations, all of which yielded realistic approximations of the reconstructed physical fields. This innovative approach represents a significant step toward an accurate characterization of the MLT dynamics, overcoming the limitations of existing methods, and providing valuable insights into the behavior of this poorly accessible region of the atmosphere. Plain Language Summary Multistatic meteor radars (MMR) have emerged as groundbreaking instruments poised to revolutionize the spatial and temporal coverage and resolution of observations in the mesosphere and lower thermosphere (MLT). In this pioneering study, we present a novel method named HYPER (HYdrodynamic Point‐wise Environment Reconstructor), which leverages sparsely sampled MMR Doppler projections to characterize the wind dynamics within the MLT. This cutting‐edge method synergizes meteor radar observations with the physics prescribed by the Navier‐Stokes equations, which govern the turbulent fluid dynamics in the MLT. To find approximate solutions to these complex equations within the observational domain, we employ a deep learning framework that aligns the neural network output with the equations of motion, ensuring the cumulative solution is tuned to the observational data. To validate our approach, we implemented virtual radar systems on the outputs of a general circulation model (GCM) and a high‐resolution direct numerical simulation (DNS) of atmospheric flow, both serving as ground‐truth 4D wind fields. Our compelling results reveal that the proposed method not only succeeds in retrieving realistic horizontal wind fields but also provides remarkably accurate estimates of vertical winds, even amidst the challenges posed by noisy Doppler observations. Key Points A method integrating physics‐informed neural networks with multistatic meteor radar detections, dubbed HYPER, to estimate 4D wind fields HYPER reproduces large‐scale and inertial range dynamics as observed in turbulent atmospheric simulations and global circulation models HYPER‐derived 4D mesoscale and mean vertical winds are physically realistic, with mesoscale vertical winds varying few m/s