Title: Assessment of Tropical Cyclone Climatology and Forecast Skill in the NASA GEOS-S2S-3 Forecast System
Abstract:
"In this study, we evaluate tropical cyclone (TC) climatology, structure, and subseasonal prediction skill in NASA’s Goddard Earth Observing System Subseasonal-to-Seasonal forecast system, Version 3 (GEOS-S2S-3), using retrospective forecasts. The model’s performance is assessed relative to observed best-track datasets (IBTrACS) and compared with the preceding version, GEOS-S2S-2. GEOS-S2S-3 produces approximately 1.5 times the observed annual TC counts across most ocean basins, whereas GEOS-S2S-2 slightly underestimates these counts. Both systems capture the observed seasonal cycle of TC activity but underestimate storm intensity, consistent with their 0.5° atmospheric resolution. GEOS-S2S-3 also simulates longer TC lifespans, which, together with its higher genesis rate, result in greater modified accumulated cyclone energy than in GEOS-S2S-2. To investigate the excessive TC genesis in GEOS-S2S-3, we examine the large-scale environment, convective precipitation, and the vertical structure of convection. Neither large-scale environmental favorability nor differences in mean convective precipitation adequately explain the genesis bias. Instead, the analysis points to stronger lower-tropospheric ascent in GEOS-S2S-3 as a potentially relevant contributor, possibly associated with changes in the tuning of moist-process parameterizations. Storm-centered diagnostics further reveal a more realistic TC vortex in GEOS-S2S-3, characterized by stronger inner-core ascent, higher mid-level relative humidity, and stronger tangential winds than in GEOS-S2S-2. Nevertheless, neither system simulates a distinct eyewall at 0.5° resolution.
The overestimation of annual TC frequency is reflected in its forecasting skill in GEOS-S2S-3 relative to GEOS-S2S-2, as assessed using Brier Skill Scores (BSS). This lack of skill in the new system primarily arises from substantial conditional biases (driven by higher variance) and unconditional biases (positive biases in mean TC genesis). A linear regression–based calibration markedly improves model skill across all basins and lead times, bringing GEOS-S2S-3 performance in line with GEOS-S2S-2 (also after calibration) and yielding skillful forecasts across all basins and lead times."