Whether tropical cyclones (TC) possess chaotic dynamics is an open question in current TC research. The existence of such chaotic dynamics is profound for TC model development and operational forecast, as it sets a limit on how much one can further improve intensity forecast skills or models in the future. Rapid advances of machine learning (ML) techniques and applications open up an opportunity to explore TC intensity chaos from a different angle. Building upon our recent results on the low-dimensional chaos of TC intensity, this study presents a novel use of ML models to quantify TC intensity chaos. By treating TC scales as input features for ML models, we show that TC intensity displays a limited predictability range of ~3 hours due to chaotic variability at the potential intensity (PI) equilibrium. This short predictability range for TC intensity is robust across ML architectures including deep neural networks (DNN), gated recurrent units (GRU), and long-short term memory (LSTM) examined in this study. Using the minimum central pressure as a metric for TC intensity could extend the predictability range to 5-6 hours, yet the limited predictability for TC intensity is still well captured in all ML models. As a result, the intrinsic variability of TC intensity related to low-dimensional chaos prevents intensity errors in any TC model from being arbitrarily reduced, regardless of how perfect a TC model or vortex initialization is. Our findings support the existence of chaotic dynamics at the PI limit and demonstrate an innovative way of applying ML to study atmospheric predictability.

This study examines the large-scale factors that govern global tropical cyclone (TC) formation and an upper bound on the annual number of TCs. Using idealized simulations for an aqua-planet tropical channel, it is shown that the tropical atmosphere has a maximum capacity in generating TCs, even under ideal environmental conditions. Regardless of how favorable the tropical environment is, the total number of TCs generated in the tropical channel possesses a consistent cap across experiments. Analyses of daily TC genesis events reveal further that global TC formation is intermittent throughout the year in a series of episodes at a 2-week frequency, with a cap of 8-10 genesis events per episode. Examination of different large-scale environmental factors shows that 600-hPa moisture content, 850-hPa absolute vorticity, and vertical wind shear are the most critical factors for this global episodic TC formation. Specifically, both the 850 hPa absolute vorticity and the 600 hPa moisture are relatively higher at the onset of TC formation episodes. Once TCs form and move to poleward, the total moisture content and the absolute vorticity in the main genesis region subside, thus reducing large-scale instability and producing an unfavorable environment for TCs to form. It takes $\sim$2 weeks for the tropical atmosphere to remoisten and rebuild the large-scale instability associated with the Inter-Tropical Convergence Zone before a new TC formation episode can occur. These results offer new insight into the processes that control the upper bound on the global number of TCs in the range of 80-100 annually.

This study examines the characteristics of tropical cyclone (TC) turbulence and its related predictability implications. Using the Fourier-Bessel spectral decomposition for convective-permitting simulations, it is shown that TC turbulence possesses different spectral properties in the azimuthal and radial directions, with a steeper power law in the radial-wavenumber than those in the azimuthal-wavenumber direction. This spectral difference between the azimuthal and radial directions prevents one from using a single wavenumber to interpret TC intensity predictability as for classical homogeneous isotropic turbulences. Analyses of spectral error growth for a high-wavenumber perturbation further confirm that the spectral growth is more rapid for high azimuthal wavenumbers than for the radial wavenumbers, reaching saturation after ~9 hrs and ~18 hrs for the azimuthal and radial directions, respectively. This result highlights the key difficulty in quantifying TC intensity predictability based on spectral upscale error growth for future applications.