Data-Driven Gravity Wave Source Parameterization Using Machine Learning

Representing gravity wave (GW) sources accurately remains a major challenge for climate models. While parameterizations for orographic and convective gravity waves are well established, studies have shown that additional sources, including fronts, jet streams, and jet exit regions, also generate gravity wave activity. These sources driven by dynamics are often not clearly defined in current parameterization methods, which leads to biases in momentum deposition and large-scale circulation.
In this study, we propose a machine learning-based framework to model gravity wave sources in a unified and data-driven way. We use high-resolution ICON simulations to resolve gravity wave generation from a wide range of atmospheric processes. A reduced-order representation of the gravity wave action density spectrum serves as the target function. This allows for a compact yet meaningful description of gravity wave emission. Input features include resolved large-scale flow characteristics, subgrid-scale orographic properties, and convective indicators taken from the model fields.
We train supervised machine learning models to learn the nonlinear relationship between the atmospheric state and the resulting gravity wave emission. The resulting parameterization accounts for gravity wave generation related not only to orography and convection but also to dynamically driven sources such as frontogenesis and jet-related processes.