Hurricane Predictability Analysis with Singular Vectors in the Multiresolution Global Shallow Water Model

Abstract

In this study, the tangent linear and adjoint (TL/AD) models for the Model for Prediction Across Scales (MPAS) Shallow Water (SW) component are tested and demonstrated. Necessary verification check procedures of TL/AD are included to ensure that the models generate correct results. The TL/AD models are applied to calculate the singular vectors (SVs) with a 48-h optimization time interval (OTI) under both the quasi-uniform-resolution (UR) and smoothly variable-resolution (VR) meshes in the cases of Hurricanes Sandy (2012) and Joaquin (2015). For the global domain, the VR mesh with 30 210 grid cells uses slightly fewer computational resources than the UR mesh with 40 962 cells. It is found that at the points before Hurricanes Sandy and Joaquin made sharp turns, the leading SV from the VR experiment show sensitivities in both areas surrounding the hurricane and those relatively far away, indicating the significant impacts from the environmental flows. The leading SVs from the UR experiments are sensitive to only areas near the storm. Forecasts by the nonlinear SW model demonstrate that in the VR experiment, Hurricane Sandy has a northwest turn similar to the case in the real world while the storm gradually disappeared in the UR experiment. In the case of Hurricane Joaquin, the nonlinear forecast with the VR mesh can generate a track similar to the best track, while the storm became falsely dissipated in the forecast with the UR mesh. These experiments demonstrate, in the context of SW dynamics with a single layer and no physics, the track forecasts in the cases of Hurricanes Sandy and Joaquin with the VR mesh are more realistic than the UR mesh. The SV analyses shed light on the key features that can have significant impacts on the forecast performances.