TOWARDS A SELF-CONTAINED ANALYSIS AND REANALYSIS SYSTEM FOR HURRICANES

Abstract

Tropical cyclone (TC) prediction is one of the most challenging problems in atmospheric science. While track forecasts have improved dramatically over the past several decades, intensity forecasting continues to lag behind, particularly in cases of rapid intensification. This disparity reflects fundamental challenges: TC intensity is governed by complex interactions between inner-core dynamics, boundary layer processes, air-sea fluxes, and broader environmental factors. Numerical models must achieve very high resolution to resolve the eyewall region while simultaneously covering a domain large enough to capture steering flow and synoptic-scale interactions. TCs develop predominantly over observation-sparse oceans and thus rely heavily on satellite measurement systems that may introduce significant biases. Modern TC-focused systems, such as NOAA's Hurricane Analysis and Forecast System (HAFS), address some of these challenges but rely heavily on global models for initial and boundary conditions, and bias correction parameters. This dependence limits our ability to diagnose deficiencies within the model itself and impedes progress toward a fully mature, self-contained, ensemble-based TC forecasting system. This dissertation advances hurricane forecasting through three complementary studies. First, we perform idealized experiments within a low-dimensional chaotic dynamical model to examine fundamental limitations of observation bias correction in the presence of model bias. We find that anchor observation-based methodologies can successfully separate biases only in proximity to anchor observations, but that without sufficient coverage, bias reinforcement and drift can occur. Ultimately, diagnosing and addressing model bias is the best strategy for mitigating bias reinforcement. Second, we demonstrate that HAFS can operate in a fully-cycled, self-contained mode with online satellite radiance bias correction. Experiments show that training bias correction coefficients on the native HAFS grid produces improved TC forecasts compared to adopting coefficients from external global models. Third, we extend the self-contained framework to a basin-scale ensemble data assimilation and forecasting system. Comparing different resolution configurations reveals substantial differences in cloud coverage and composition, generating significant domain-wide biases with consequent impacts on TC forecasts. Furthermore, we demonstrate the benefit of an ensemble-based framework. While obstacles remain, clear patterns emerge that inform future model development. Taken together, this work establishes a foundation for a self-contained, continuously-cycled, probabilistic TC analysis system, essential for diagnosing model deficiencies and for developing an analysis of record to support forecast verification and machine learning applications.