Supplementary material to "Evaluation of North Atlantic Tropical Cyclones in a Convection-Permitting Regional Climate Simulation"

Lara Tobias-Tarsh, Chunyong Jung, Jiali Wang, Vishal Bobde, Akintomide A. Akinsanola, V. R. Kotamarthi2025 · DOI

Abstract

This research evaluates the performance of a high-resolution, convection-permitting regional climate simulation (ADDA_v2) in capturing the behavior of North Atlantic Tropical Cyclones (TCs). By comparing the simulation against ERA5 reanalysis and the HURDAT2 observational record, the study assesses track density, intensity distributions, and spatial patterns across multiple resolutions. The results indicate that while ADDA_v2 significantly improves upon the frequency biases found in ERA5, it introduces a tendency to over-produce storm counts and remains statistically distinct from historical observations.

Introduction

Tropical Cyclones are among the most destructive weather events on Earth, yet they are notoriously difficult to simulate in climate models. Standard models often lack the resolution to capture "convection-permitting" scales—the small-scale, intense updrafts and cloud processes that drive a hurricane's core. This study addresses the gap between coarse-resolution global models and the need for high-fidelity local data by evaluating whether a convection-permitting scale produces more realistic storm tracks, frequencies, and physical characteristics.

Methods

The researchers employed a multi-scale statistical approach to evaluate the simulation:

  • Track Density Analysis: TC centroids were calculated within specific grid cells over a 20-year period. To ensure results were not artifacts of the grid, three resolutions were tested: 2.5°, 1.5°, and 1.0° bins.
  • Statistical Metrics: Performance was measured using Root Mean Square Error (RMSE), Pattern Correlation (spatial accuracy), and "Modeling Yield" (the ratio of simulated to observed storm counts).
  • Distribution Testing: Non-parametric tests, including Kolmogorov-Smirnov (K-S), Cramér-von Mises, and Anderson-Darling, were used to compare the entire shape of wind speed and pressure distributions against observations.
Figure 1: Comparison of TC track density across ERA5, ADDA_v2, and HURDAT2 at varying spatial resolutions.

Key Results and Discussion

Track Density and Modeling Yield

A primary finding is the difference in "Modeling Yield." ERA5 consistently under-produces TC track density with a yield of approximately 0.60. In contrast, the ADDA_v2 simulation tends to over-produce them, showing a yield of ~1.27 at the 2.5° resolution. This suggests that while high-resolution models solve the "missing storm" problem inherent in coarser models, they may introduce a "too many storms" bias.

At a 2.5° resolution, ADDA_v2 achieved an RMSE of 9.88, outperforming ERA5’s 11.19. Interestingly, ERA5 often shows a higher Pattern Correlation (up to 0.92) than ADDA_v2 (0.84), indicating that while ERA5 places storms in the correct locations more consistently, ADDA_v2 captures the local intensity of track density more accurately in certain regions.

Table S1: Summary of statistics for track density at each bin size, highlighting the performance of ERA5 vs. ADDA_v2.

Resolution Sensitivity and Intensity

As the analysis grid becomes finer (from 2.5° down to 1.0°), the RMSE drops significantly for both models. For physical characteristics like 10m wind speeds and Minimum Sea Level Pressure (MSLP), ADDA_v2 showed lower test statistics in the K-S and Anderson-Darling tests compared to ERA5. This indicates that ADDA_v2's distribution of storm intensities is closer to the observed HURDAT2 data than the reanalysis is.

Conclusions and Limitations

Despite the improvements offered by the convection-permitting simulation, p-values for most distribution tests remained below 0.01, meaning the differences between the models and observations are still statistically significant. The high "Modeling Yield" in ADDA_v2 suggests the model may be over-simulating weak systems or failing to properly decay storms in hostile environments. Future work will focus on tuning convection-permitting parameters to reconcile these frequency biases with observed storm life cycles.

Summary updated

Abstract

Precision Difference Mean Error Modeling Yield Pattern Correlation Q95 diff 2.5 bin ERA5 11.19 -6.17 -6.95 0.60 0.92 -16.0 ADDA_v2 9.88 0.76 0.24 1.27 0.84 -1.0 1.5 bin ERA5 4.38 -2.31 -2.38 0.69 0.88 -6.0 ADDA_v2 4.78 0.31 0.11 1.18 0.76 0 1.0 bin ERA5 2.32 -1.04 -1.10 0.76 0.81 -3.0 ADDA_v2 2.76 0.13 0.05 1.38 0.67 0 Table S1.Summary of statistics for track density at each bin size, as shown in Fig. S1.RMSE Precision Difference Mean Error Modeling Yield Pattern Correlation Q95 diff Track Density ERA5 11.31 -6.17 -6.95 0.60 0.90 -16.0 ADDA_v2 9.92 0.76 0.24 1.27 0.84 -1.0 Initial Location ERA5 1.17 -0.78 -0.28 0.64 0.83 0.0 ADDA_v2 1.34 -0.39 -0.003 0.57 0.70 1.0 Translation Speed ERA5 3.89 0.31 0.56 0.73 0.82 N/A ADDA_v2 4.59 0.08 1.21 1.03 0.72 N/A Table S2.As in Table S1, but for the track density, initial location, and translation speed using 2.5 bins, as shown in Fig. 3.
BibTeX
@misc{tobiastarsh2025supplementary,
  author = {Lara Tobias-Tarsh and Chunyong Jung and Jiali Wang and Vishal Bobde and Akintomide A. Akinsanola and V. R. Kotamarthi},
  title = {{Supplementary material to "Evaluation of North Atlantic Tropical Cyclones in a Convection-Permitting Regional Climate Simulation"}},
  year = {2025},
  doi = {10.5194/egusphere-2025-1805-supplement},
}
Figure