Convection-Generated Gravity Waves from Weather Radar Observations Via a Neural Network and a Dynamical Atmospheric Model

Convection-generated gravity waves (CGW) transport momentum and energy vertically from weather systems upward to the stratosphere where they drive global features of Earth’s circulation like the tropical Quasi-Biennial Oscillation. The gravity waves are subgridscale processes in global climate and weather forecasting models, and CGW drag on the circulation is a parameterized process used to reduce stratospheric wind biases and improve global weather teleconnections that have stratospheric pathways. While convection-permitting or -resolving simulations may include CGW, validation of the waves is rarely possible as the waves are highly transient and the simulations cannot reproduce the CGW forcing at the correct times, locations, and/or intensities.

Here we derive the CGW forcing from weather radar observations using two methods: Neural networks and a previously developed look-up table method. Both methods are trained on full-physics simulations performed with the Weather Research and Forecasting (WRF) model to learn relationships between radar-observed properties of precipitation and the 3-dimensional time-dependent structure of convective latent heating, which is the source of CGW. The results show sensitivity to the type of convection simulated/range of training data, and we discuss similarities/differences between the two methods. The heating rates are then used to force an idealized GW-resolving dynamical model. Simulated CGW forced in this way closely resemble satellite observations of waves in the stratosphere. CGW drag in these validated simulations extends 100s of kilometers away from the convective sources, highlighting errors in current gravity wave drag parameterizations due to the use of the single-column physics approximation. Such validated simulations have significant potential to further basic understanding of CGW, improve their parameterization physically, and provide important constraints on parameterization tuning for global climate and weather models.