Prediction and Understanding of Subgrid-Scale Vertical Fluxes by Missing Midlatitude Mesoscale Processes Using a Machine Learning Approach

Mesoscale atmospheric processes that are neither resolved nor parameterized in global climate models, such as slantwise convection, can have a significant impact on climate variability and change. An example of such mesoscale influence on the climate is shown by Wills et al. 2024,  who demonstrate that circulation responses to surface anomalies are increased through heat and momentum fluxes by mesoscale processes. To enable longer simulations in comprehensive global climate models that include information about subgrid mesoscale processes, a machine learning (ML) parameterization could be applied at relatively low computational cost. So far, such ML parameterizations have been primarily applied to idealized geographies (e.g., aquaplanets), and they have not targeted midlatitude mesoscale processes in particular.

In this work, we focus on midlatitude mesoscale processes over the Gulf Stream region, as simulated by variable resolution CESM2 simulations, which have 14-km resolution over the North Atlantic. Learning subgrid fluxes from this model allows a targeted parametrization of mesoscale processes leading to vertical fluxes, namely slantwise convection and frontogenesis. We use an artificial neural network to predict vertical profiles of subgrid fluxes of momentum, heat and moisture. The features (inputs) for the ML models in this work include coarse-grained atmospheric state variables at each grid point, such as the vertical profiles of horizontal winds, temperature and their horizontal shear as well as surface pressure. The vertical profile of the specific humidity and the value of convective available potential energy are included to assess the importance of moist dynamics in the determination of subgrid convectional fluxes. Our results show that moisture variables have a rather small impact, suggesting that the subgrid fluxes can be explained by dry dynamics. A greater importance is found in the horizontal differences of neighbouring momentum and temperature columns. This suggests that neighbouring column information may be essential in the prediction of subgrid-scale fluxes, e.g., through the action of shear instabilities or conditional symmetric instability. Combined with information about the vertical localization relationship of the inputs and outputs, the goal is to feed this information into an equation discovery approach, which could lead to deeper physical understanding of mesoscale momentum and energy fluxes in midlatitudes.