Forecasting Aviation Turbulence in Tropical Airspace: A Comparative Evaluation of Machine Learning and Deep Learning Models in Jakarta FIR

Atmospheric turbulence, particularly Convectively Induced Turbulence (CIT), poses a significant hazard to aviation safety within the Jakarta Flight Information Region (FIR), a region characterized by persistent deep convection over the Indonesian Maritime Continent. While machine learning models exhibit high proficiency in forecasting clear air turbulence (CAT) in mid-latitude regions, their predictive capability for CIT in tropical maritime environments remains underexplored. In this study, we evaluated four AI architectures for forecasting tropical turbulence, which are eXtreme Gradient Boosting (XGBoost), Artificial Neural Networks (ANNs), Bidirectional Long Short-Term Memory (Bi-LSTM), and Gated Recurrent Units (GRUs). Models were trained using fifteen dynamic and thermodynamic atmospheric predictors derived from ECMWF ERA5 reanalysis data and validated against operationally verified Pilot Reports (PIREPs) from January 2024 to April 2026. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), while operational performance was assessed through threshold optimization (Default, Youden Index, F1-score) using the Probability of Detection (POD) and the Matthews Correlation Coefficient (MCC). XGBoost demonstrated superior discrimination capability (AUC = 0.65), outperforming ANN (0.60), GRU (0.60), and Bi-LSTM (0.59). These relatively modest metrics underscore the inherent challenge of resolving microscale, rapidly evolving CIT using coarse resolution reanalysis datasets. Nevertheless, threshold optimization via the Youden Index significantly enhanced detection accuracy; XGBoost achieved the most balanced operational efficacy (MCC ≈ 0.26, False Alarm Ratio ≈ 0.52, POD ≈ 0.62), while both XGBoost and Bi-LSTM optimized the detection of moderate-to-severe events. These findings elucidate the fundamental differences in turbulence predictability between tropical convective environments and CAT-dominated regimes, substantiating the necessity for region specific, AI-driven meteorological frameworks to advance next-generation aviation warning systems.