Mid-level longwave heating accelerates tropical cyclogenesis: insights from ML-informed WRF simulations

Evidence from idealized and real-case modeling suggests that the longwave (LW) component of the cloud-radiative feedback (CRF) is fundamental to tropical cyclogenesis (TCG), accelerating a ~5-day process by multiple days in idealized studies. However, existing CRF mechanism-denial frameworks - such as spatial homogenization of radiative heating or utilizing “cloud-transparent” radiative schemes - preclude counterfactual analysis on how TCG efficiency is affected by specific radial and vertical structures of CRF, which is determined by the distribution of different cloud types in the TC. 

 

To address whether an “optimal LW CRF pattern” that maximizes TCG efficiency exists, we developed an ML-informed WRF modeling framework that enables counterfactual experiments by adding an external 3D, ML-discovered, heat forcing to the total heating tendency returned by the RRTMG longwave radiation scheme. Physics-informed inverted LASSO regressions, trained on a WRF ensemble on Typhoon Haiyan (2013), isolate an “optimal LW perturbation” in the form of inner-core mid-level heating. which is strikingly different from a longwave perturbation that maximizes near the cloud top obtained with simple data analysis (2-day azimuthal mean).

 

We conduct a series of pattern-perturbation experiments to validate this data-driven proposed “optimal CRF pattern”: the ML-discovered mid-level perturbation accelerates the intensification of Haiyan more efficiently than an empirical 2-day mean upper-level LW perturbation. Changes in vertical velocity precede changes in precipitation characteristics in the perturbation experiments, establishing a causal chain from CRF to intensity change. The mid-level-heating runs exhibit higher inner-core stratiform fractions, more intense convective bursts, stronger mid-level vorticity, and lower mean sea-level pressure. These results demonstrate that ML can serve as an objective hypothesis generator, reducing the scientific search space and facilitating efficient data-driven discovery of the structural drivers to TCG.