Identifying Three-Dimensional Radiative Patterns Associated with Early Tropical Cyclone Intensification

Abstract

Cloud radiative feedback impacts early tropical cyclone (TC) intensification, but limitations in existing diagnostic frameworks make them unsuitable for studying asymmetric or transient radiative heating. We propose a linear Variational Encoder-Decoder (VED) to learn the hidden relationship between radiation and the surface intensification of realistic simulated TCs. Limiting VED model inputs enables using its uncertainty to identify periods when radiation has more importance for intensification. A close examination of the extracted 3D radiative structures suggests that longwave radiative forcing from inner core deep convection and shallow clouds both contribute to intensification, with the deep convection having the most impact overall. We find that deep convection downwind of the shallow clouds is critical to the intensification of Haiyan. Our work demonstrates that machine learning can discover thermodynamic-kinematic relationships without relying on axisymmetric or deterministic assumptions, paving the way towards the objective discovery of processes leading to TC intensification in realistic conditions.

Figures (9)

• Figure 1: The interpretable linear VED framework proposed in this study combines a pattern-finding encoder and a decoder for TC intensification rate prediction. The first linear layer in the encoder modules combines different radiation structural information in the PCs into the time evolutions of the mean structures and uncertainty structures. A random sampling of the normal distributions with mean and log variances conditioned on the inputs introduces uncertainty to a decoder module that predicts the 24-hour surface wind intensification prediction. There is flexibility in the choice of the decoder architecture, from simple multiple linear regressors to a complex nonlinear Artificial Neural Network (ANN). Optimization of model weights with loss functions tries to minimize the absolute error between the truth and the predictions from the decoder based on the learned patterns.
• Figure 2: The mean CRPS scores (left column) and the spread-skill diagrams (right column) show that the trained VED models outperform the trained baseline models with different dropout rates and degrees of nonlinearity in the prediction equation ("nonln:3" represents baseline model with three nonlinear layers in the decoder). Comparing the best fully linear baseline model for Haiyan and Maria (brown lines in panel b and d; $\mathbf{dropout\ rates}$ of 0.3 and 0.1 respectively) and the mean performance of the VED model with the best SSREL score shows that the VED model makes fewer mistakes in its predictions and is generally more well-calibrated than the best baseline model. A well-calibrated model means that most points on the model's spread-skill curve are as close to the 1-1 line (gray lines in b, d) as possible.
• Figure 3: Decomposing tropical cyclone intensification predictions for Maria and Haiyan shows longwave radiative heating's link to early intensity differences. We present mean VED surface intensification predictions (dashed) for Maria (a) from three WRF simulations and for Haiyan (c) from two ensemble members, with actual rates of intensity change (thick). The shadings in the left columns illustrate the range of possible VED predictions given the same inputs using a Monte Carlo approach. Panels (b) and (d) show longwave and shortwave radiation contributions to the mean VED predictions. Zooming in on Haiyan Member 2, Panel (e) compares the evolution of VED uncertainty associated with longwave radiation (purple) against two MSE variance sources (blue and brown). In Maria's mechanism-denial experiment and the early phase of Haiyan, the model associates the difference between a quickly intensifying TC and a slowly intensifying one with longwave radiation. The vertical dashed lines in panel (e) show the two-stage behavior in the longwave MSE variance source term (black) and similar behavior in VED prediction uncertainties (purple). The time lag between the two methods is shown in panel (e; black arrow).
• Figure 4: The learned data-driven structure for the mean longwave prediction ($\Pi_{\mu\mathrm{LW}}$; c) for Maria shows a prominent upper-level longwave anomaly dipole. A sample from Hour 80 of the CTRL simulation (a-b), which is predicted by the ML model to have a high intensification rate because of a positive projection of $LW^\prime$ (b; perturbation compared to the training mean) onto $\Pi_{\mu\mathrm{LW}}$ (c). We contrast the CTRL sample with another sample taken at the same hour from the NCRF-36h simulation (d-e), which is predicted to have a negative intensification rate due to a negative projection of $LW'$ (e) onto $\Pi_{\mu\mathrm{LW}}$. The two samples illustrate the physical meaning of the projections: a positive projection (red arrows) occurs when the $LW'$ is similarly spatially to $\Pi_{\mu\mathrm{LW}}$, whereas a negative projection occurs when the $LW'$ is opposite in sign to $\Pi_{\mu\mathrm{LW}}$.
• Figure 5: The learned data-driven structure for the mean longwave prediction ($\Pi_{\mu\mathrm{LW}}$; c) for Haiyan shows the spatial distribution of $LW^\prime$ that is most correlated with early TC intensification. The example that is predicted to have a high intensification rate (a,b) is taken from Hour 15 of Member 2, whereas the predicted low intensification rate (d,e) example is taken from Hour 17 of Member 11. The Member 2 $LW^\prime$ example has a strong positive projection (red arrows) onto $\Pi_{\mu\mathrm{LW}}$ (c), which is indicative of concentrated inner-core deep convection (200 hPa longwave anomaly dipole within 100 km of TC center) in the raw azimuthal-averaged $\mathrm{LW}$. Additionally, the Member 2 example also features shallow clouds in the outer core (900 hPa anomaly dipole between 100-300 km from the TC center; b). Both the inner-core deep convection and outer-core shallow cloud signatures are weaker in the Member 11 example, which leads to a lower predicted intensification rate (blue arrow).