Using Machine Learning to Estimate Nonorographic Gravity Wave Characteristics at Source Levels

Machine learning (ML) provides a powerful tool for investigating the relationship between the large-scale flow and unresolved processes, which need to be parameterized in climate models. The current work explores the performance of the random forest regressor (RF) as a nonparametric model in the reconstruction of nonorographic gravity waves (GWs) over midlatitude oceanic areas. The ERA5 dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF) model outputs is employed in its full resolution to derive GW variations in the lower stratosphere. Coarse-grained variables in a column-based configuration of the atmosphere are used to reconstruct the GWs variability at the target level. The first important outcome is the relative success in reconstructing the GW signal (coefficient of determination R² ’ 0.85 for “E3” combination). The second outcome is that the most informative explanatory variable is the local background wind speed. This questions the traditional framework of gravity wave parameterizations, for which, at these heights, one would expect more sensitivity to sources below than to local flow. Finally, to test the efficiency of a relatively simple, parametric statistical model, the efficiency of linear regression was compared to that of random forests with a restricted set of only five explanatory variables. Results were poor. Increasing the number of input variables to 15 hardly changes the performance of the linear regression (R² changes slightly from 0.18 to 0.21), while it leads to better results with the random forests (R² increases from 0.29 to 0.37).