Assessing the Impact of High-Frequency Data on Statistical-Dynamical Downscaled Tropical Cyclones

Statistical-dynamical downscaling (SDD) tropical cyclone (TC) models utilize a climate model’s large-scale climatology to generate synthetic storms. While SDD methods can improve on some aspects of GCM depiction of TC climatology, they have also been shown to inherit model biases, generate unrealistic track patterns, and overrepresent landfalling storms (Bolivar and Zarzycki 2024, submitted to Journal of Climate). This has important implications for stakeholders who use SDD models to inform risk-related decision-making. Currently, SDD models use monthly mean thermodynamic fields to predict TC intensification and daily kinematic fields to calculate monthly variances used to generate stochastic winds along the storm track. This approach is inexpensive but has two key disadvantages: 1) temporal information is limited to the month of occurrence, and 2) storms are steered by stochastic instead of “true” winds, which produce abnormalities in tracks. Using sub-monthly input data for SDD models may ameliorate some of these problems and expand the applicability of SDD models beyond providing high-level risk assessment.

Using a method based on Davis et al. (2008), we remove irrotational and nondivergent winds and use a sigmoid smoothing function from HighResMIP data to create vortex-removed 6-hourly wind input data. We leverage the SDD TC model described in Lin et al. (2023) to generate tracks using these data. As proof of concept, we use manual seeding to test whether storm tracks already existing in GCM can be reasonably reproduced. A full set of randomly seeded tracks is also generated and compared to observed landfalls using IBTrACS and objectively tracked storms in reanalyses, as well as objectively tracked storms in HighResMIP models and synthetic storms generated using monthly wind input data. These results suggest possible pathways for improving SDD models and help shed light on the direct simulation of TCs in climate models.