Abstract Clear-air turbulence (CAT) remains one of the most challenging hazards in aviation meteorology due to its abrupt onset, limited observability, and the well-known limitations of conventional diagnostic indices. This study applies machine-learning (ML) algorithms to improve CAT prediction over southern Brazil using atmospheric predictors derived from the Global Forecast System (GFS) model. Fourteen physically based variables, representing wind shear, thermodynamic stability, deformation, and classical turbulence diagnostics, were extracted from GFS outputs between 500 and 200 hPa and used to train and evaluate fifteen ML classifiers. Model development followed a structured workflow comprising stratified sampling, feature engineering, hyperparameter optimization, and rigorous validation using robust metrics, including the Probability of Detection for CAT (PODy) and non-CAT cases (PODn), the True Skill Statistic (TSS), and the Area Under the ROC Curve (AUC). Ensemble tree-based models showed the highest skill, with the Bagging classifier achieving the best overall performance (AUC = 0.86, TSS = 0.55). Feature-exclusion analysis revealed that potential temperature (Θ) and shear-related predictors—vertical wind shear (VWS) and differential wind speed between levels (DWS), were the most influential variables, whereas turbulent kinetic energy (TKE), strongly smoothed at the GFS 0.25° resolution, contributed minimally. The results demonstrate that regionally trained ML models substantially outperform individual CAT diagnostics, capturing the multi-mechanism nature of turbulence generation along the subtropical jet over southern Brazil. From an operational standpoint, the demonstrated skill of the Bagging classifier indicates strong potential for improving early CAT alerts in Brazilian airspace, reducing reliance on subjective aircraft reports and enhancing aviation safety along one of South America’s busiest flight corridors.
Mello et al. (Mon,) studied this question.