Modulation of tropical cyclogenesis on subseasonal-to-interannual timescales in the deep-learning climate emulator ACE2

Mu-Ting Chien , Elizabeth A Barnes, Eric D Maloney

Deep-learning global climate emulators are providing a new lens to investigate tropical cyclogenesis (TC genesis). However, without explicitly enforcing known physics, it is necessary to assess whether TC genesis in these models is physical. To address this question, we use the Ai2 Climate Emulator version 2 (ACE2) trained on ERA5 reanalysis to investigate TC genesis and its relationship with the large-scale environment on subseasonal-to-interannual timescales. We run simulations with ACE2 using forcing fields from 2001 to 2010, which is outside of its training period. Compared to observations, the geographic distribution and annual cycle of TC genesis are reasonably represented in ACE2 across the globe. TC genesis in ACE2 generally occurs with favorable environmental conditions (high genesis potential index) on annual, interannual, and subseasonal timescales. On subseasonal timescales, ACE2 shows that the environmental conditions for TC genesis are affected by the occurrence of the Madden-Julian Oscillation (MJO) and convectively coupled equatorial waves (CCEWs), as in observations. With pronounced eastward propagation of the MJO and realistic simulation of precipitation and three-dimensional circulation anomalies in ACE2, a clear signal of MJO modulation of TC genesis is found in most basins. This study suggests that deep learning climate emulators may be a useful tool for understanding cyclogenesis in the current climate from subseasonal-to-interannual timescales, as well as their changes in altered climates.