FAST-ML: Hybrid Physics–Machine Learning for Tropical Cyclone Intensity Forecasting

Physics → Learned Correction → Improved Forecast → Probabilistic Ensemble → Extreme-Event Risk

FAST-ML is a hybrid physics–machine learning framework that improves tropical cyclone intensity prediction while preserving physical structure and enabling probabilistic forecasting. Rather than replacing a physical intensity model with a purely data-driven predictor, FAST-ML learns systematic corrections to the physical forecast using large-scale atmospheric information.

Graduate research at the Climate Risk & Extreme Dynamics Lab, Georgia Tech, advised by Prof. Ali Sarhadi. Manuscript submitted to Journal of Advances in Modeling Earth Systems (JAMES); available as arXiv:2609.25505 (Xiao et al., 2026).

Atmospheric State → Environmental Feature Extraction → Physics Model + ML Correction → FAST-ML Forecast → Probabilistic Ensemble → Extreme-Event / Tail-Risk Analysis

Key Result — Held-Out 2024 North Atlantic Evaluation

On the held-out 2024 North Atlantic hurricane season, FAST-ML reduced mean intensity RMSE from 15.42 kt to 10.78 kt, corresponding to a 30.1% reduction relative to the physics baseline. FAST-ML improved forecasts for 10 of 12 evaluated storms.

Model	2024 RMSE
Physics baseline (FAST)	15.42 kt
FAST-ML	10.78 kt
Relative reduction	30.1%

FAST-ML architecture: a two-stream CNN learns environmental diagnostics from multi-year, hourly, ~0.25° atmospheric fields (ERA5), which drive a mechanistic intensity ODE with learned corrections. All temporal evolution is governed by the physics.

Why FAST-ML?

Tropical cyclone intensity prediction is difficult because storm evolution depends on nonlinear interactions between environmental forcing, internal storm dynamics, and uncertain initial conditions. Purely data-driven predictors discard physical structure; purely physical models carry systematic biases. FAST-ML keeps the mechanistic model in the loop and learns where it is wrong, from the environment in which the storm evolves.

FAST-ML is part of a broader effort toward hybrid scientific machine learning for nonlinear physical systems. The goal is not to replace physical modeling, but to combine physical structure with modern machine learning for improved prediction, uncertainty quantification, and extreme-event characterization.

Main Contributions

Hybrid physics–ML forecasting — preserve mechanistic structure while learning systematic forecast discrepancies from atmospheric conditions.
Large-scale atmospheric learning — the system operates on multi-year, hourly, ~0.25° atmospheric datasets with a substantial scientific data-processing pipeline (ERA5, IBTrACS, GEFS).
Probabilistic prediction — the deterministic hybrid forecast is extended to ensembles/distributions rather than only point estimates; retaining the ODE means the spread is generated by physics, preserving the right-skewed tail.
Extreme-event and tail analysis — probabilistic forecasts allow analysis of low-probability, high-impact intensity outcomes, bridging physical forecasting and risk estimation.

Per-Storm Held-Out Evaluation

Per-storm RMSE evaluation (panel D: held-out 2024 North Atlantic season). Seven of nineteen 2024 storms are excluded by evaluation criteria (never reaching the 45 kt forecast threshold, or under 72 h of forecast after it).

Storm	Obs peak	FAST	FAST-ML	Gain
BERYL	145	12.78	10.35	+2.42
DEBBY	70	18.60	16.32	+2.28
ERNESTO	85	10.31	10.68	−0.37
FRANCINE	90	4.96	2.92	+2.04
HELENE	120	9.02	6.85	+2.17
ISAAC	90	28.97	9.17	+19.80
KIRK	130	20.59	14.90	+5.69
LESLIE	90	21.78	14.11	+7.67
MILTON	155	29.24	19.81	+9.43
OSCAR	75	8.40	5.70	+2.70
PATTY	55	3.21	6.03	−2.81
RAFAEL	105	17.20	12.52	+4.68
mean		15.42	10.78	+4.64

Probabilistic Forecasting and Uncertainty

A GEFS-based synthetic track ensemble drives FAST and FAST-ML member by member, giving a full intensity distribution to compare against the Google WeatherLab FNV3 ensemble and IBTrACS. On Hurricane Flossie (East Pacific, June 2025), 31 raw GEFS members expand to a 492-member ODE ensemble over the 5-day window:

Flossie (EP06 2025), initialised 2025-06-29 12:00 UTC: 492-member FAST-ML intensity ensemble vs the 50-member FNV3 ensemble and the best track. Left: intensity ensemble with mean and top decile. Right: per-member ventilation index.

The same ensemble feeds probabilistic products: the probability that the storm centre passes within 75 km of a point over a 5-day window, and the probability that modelled wind speed exceeds 34/50/64/83/96 kt.

Flossie: 75-km strike probability of the 492-member FAST-ML ensemble, with 31 raw GEFS members as spaghetti and the IBTrACS best track on top.

Flossie: probability of modelled wind speed exceeding 34/50/64/83/96 kt, each panel zoomed to the track domain.

Evaluation Regimes

Held-out evaluation. The 2024 North Atlantic season is the main held-out result (training and normalization use 2003–2022 data): RMSE 15.42 → 10.78 kt (−30.1%), improved 10 of 12 storms.
Historical diagnostics. Across all 224 North Atlantic storms of 2003–2024 meeting the same criteria, FAST averages 18.20 kt and FAST-ML 13.87 kt (−4.33 kt), better on 179 of 224 — reported as aggregate/historical diagnostics, not a fully independent test.
Prospective case studies. FAST-ML is additionally evaluated against Google WeatherLab FNV3 ensembles on selected 2025 tropical cyclone case studies (Flossie, Priscilla) to examine peak-intensity estimates, ensemble behavior, and physically meaningful tail outcomes.

Resources

Paper: arXiv:2609.25505
Code and data: FAST_ML (inference-only release: weights, data pipeline, evaluation, and figure reproduction)
Related: TCG (physics-guided tropical cyclone genesis prediction), FHLO (physics intensity model reference)

Data sources: IBTrACS v04 best-track data (NOAA NCEI), ERA5 reanalysis (Copernicus), GEFS reforecasts (NOAA), Google WeatherLab FNV3 ensemble forecasts.