Deep Learning-Based Reconstruction of Nonorographic Gravity Wave Patterns in the Lower Stratosphere

Recently, there has been a significant interest in applying machine learning (ML) to improve the performance of general circulation models (GCMs). Subgrid processes not resolved directly in weather and climate models still require to be parameterized. ML constitutes a set of promising methods to address the problems such as computational cost and uncertainty introduced by parameterization in numerical simulations.

The current study examines the performance of deep learning in reconstructing nonorographic gravity waves (GWs) over midlatitude oceanic regions. A convolutional neural network (CNN) is employed to predict high-resolution variables—standard deviation of momentum flux, horizontal divergence, and vertical velocity—reflecting GW activity in the lowermost stratosphere. Both the targets and the coarse-resolution explanatory variables, spanning the troposphere, are obtained from the ERA5 dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), as outlined by Amiramjadi et al. (2023).

The results demonstrate that the model effectively reconstructs the GW signal and captures the seasonal cycle of GW activity with a reasonable computational cost. The mean coefficient of determination (R²) and Pearson’s correlation coefficient (R) across all grid points in the study area are approximately 0.42 and 0.67, respectively, using all predictors.

 

Reference:

Amiramjadi, M., Plougonven, R., Mohebalhojeh, A. R., & Mirzaei, M. (2023). Using machine learning to estimate nonorographic gravity wave characteristics at source levels. Journal of the Atmospheric Sciences, 80(2), 419–440.