Extracting Three-dimensional Radiative Patterns Favoring Tropical Cyclogenesis with an Interpretable, Data-Driven Model

One of the most elusive aspects of tropical cyclone (TC) intensity evolution relates to the genesis of TCs, which involves multiscale interactions between TC circulation and mesoscale convective systems. Cloud radiative feedback has been shown to have a first-order effect on genesis, which is usually diagnosed through the moist static energy (MSE) budget or the Sawyer-Eliassen Equation (SEQ). However, these methods have important limitations - the MSE budget does not directly yield predictions on surface intensification, whereas the SEQ can only diagnose the circulation response to an axisymmetric heat forcing. The lack of spatial context, in particular, makes the current diagnostic frameworks suboptimal in (1) studying TC genesis pathways in realistic environments, and (2) clarifying the role of asymmetric radiative forcing on TC genesis.

To quantify the role of asymmetric radiative forcing on TC genesis, we developed an interpretable, variational Encoder-Decoder (VED) framework that extracts 3D radiative heating patterns informative to the probabilistic prediction of the 24-hour intensification of TC surface winds.

Our work focuses on how the VED model can be used to extract new physical insights. Applying the model to a set of convective-permitting WRF ensemble simulations on Typhoon Haiyan (2013) and Hurricane Maria (2017), the bias and uncertainties in the probabilistic VED predictions differentiate temporal periods dominated by radiation from those that are dominated by non-radiative effects. The longwave radiation contributes the most to the intensity changes for periods dominated by radiation, namely the genesis phase. The extracted 3D longwave pattern informs how radiative heating anomalies relate to surface winds. Examining the time evolution of these structures further relates these anomalies to the evolution of clouds in both TCs. The results on the Haiyan ensemble show that the radiative effects of active deep convection and shallow cumulus clouds in the TC inner core result in higher intensification rates in some members than the members that lack those features. All extracted structures show prominent wavenumber-1 asymmetry in the inner core (with positive heating anomaly downshear), suggesting that asymmetric radiative forcing is more relevant to genesis than axisymmetric ones.

To conclude, our results show that modern machine learning techniques can be a useful way to discover physically reasonable mappings between thermodynamic forcing and kinematic changes without relying on axisymmetric or deterministic assumptions. This opens the door to objective discovery of processes leading to TC genesis in realistic environments.