Explicit representation of inertia–gravity waves in general circulation models is not yet feasible due to their small scale relative to the spatial resolution of the models. With enhancement of technology in recent decades, there has been a significant interest in applying machine learning (ML) to improve the performance of general circulation models (GCMs). This research investigates the performance of Convolutional Neural Network (CNN) as a non-parametric model in reconstructing nonorographic gravity waves in mid-latitude oceanic regions using the ERA5 reanalysis data. This data set with low-resolution of 2.5° × 2.5° for predictor variables and with high-resolution of 0.25° × 0.25° for predictand variables have been selected to quantify the characteristics of gravity waves. The explanatory variables employed at 15 pressure levels are representative of gravity waves sources. These variables include temperature gradient, potential vorticity, potential vorticity anomaly, relative vorticity, pressure vertical velocity, and horizontal wind velocity. On the other hand, for the gravity-wave activity in the lowermost stratosphere, the target variables have been chosen as the standard deviation of three key parameters: horizontal divergence, absolute momentum flux, and vertical velocity at the 100 hPa pressure level. Two approaches have been evaluated; the first approach examined the effect of training data volume while keeping the number of input channels constant at 150 channels, and the second one emphasizes optimizing the number of input channels while keeping the training period constant at three years. The second approach employed three distinct configurations characterized by varying numbers of input channels: 150 channels encompassing all explanatory variables, 15 channels restricted to horizontal wind velocity only, and 42 channels selected based on the most relevant explanatory variables in reconstructing gravity waves. Data from three years—2017, 2018, and 2020—were selected for the training and validation datasets, while data from 2019 were employed for the testing dataset.
The results obtained from predicting the standard deviation of three target variables show that increasing the volume of training data in first approach is able to preserve the spatial structure and intensity of variables more effectively. Examining the results of the second approach showed that the model has low performance using 15 input channels, yet has comparable results with 42 and 150 input channels in expressing the importance of horizontal wind variable to predict target variables. The relatively similar results in the 42- and 150-channel combinations indicate the inherent capability of CNN in identifying less-important variables using appropriate activation function. Despite limitations in predicting extreme events, CNN can detect periods of gravity waves activity properly and potentially reconstructs inertia–gravity waves well. Despite the advantages of using CNN in gravity wave reconstruction, conducting more comprehensive experiments could prepare CNN for operational implementation within climate models.