Vortex center finding — Barnes analysis
The GFDL Vortex Tracker locates tropical cyclone centers with Barnes analysis:
a Gaussian distance-weighted average (e-folding radius ~60–75 km) applied to
MSLP minima and 700/850 hPa vorticity and geopotential fields. The analysis is
iterated (~4 passes) while shrinking the grid, reaching about 1/16° positional
precision even from coarse model output, and the per-variable centers are
averaged into a consensus fix.
Barnes analysis
Gaussian weighting
Iterative refinement
Multi-variable consensus
Ensemble prediction with Bred Vectors
Deterministic AI weather models are extended into ensembles by applying Bred
Vector initial perturbations — fast-growing error modes bred from repeated
short forecasts — to evaluate the predictability of tropical cyclone tracks
over the Western North Pacific. GenCast-based probabilistic ensembles are
compared alongside.
Bred Vector
Initial perturbations
GenCast ensembles
Intensity calibration
AI models systematically underestimate tropical cyclone intensity. Statistical
bias correction is compared with XGBoost regression for central pressure and
maximum wind speed over the Western and Eastern Pacific basins, with careful
handling of units (hPa, knots), coordinate conventions, and missing-value
semantics.
Bias correction
XGBoost
WP · EP basins
Verification
Forecast tracks and intensities are verified against KMA forecast records and
the IBTrACS best-track archive — track error by lead time, MSLP and MWS error
distributions — with dataset quality controlled by cross-validating initial
times and storm identifiers before scoring.
IBTrACS
Track error by lead time
MSLP · MWS errors