Machine learning application and operational strategy for global low-level aviation turbulence forecasting

Abstract Low-level turbulence (LLT), primarily driven by terrain-induced and convective processes, remains a critical hazard to aviation safety. This study establishes the applicability of machine-learning to global LLT forecasting below 10,000 ft, alongside the LLT-adapted Graphical Turbulence Guidance (GTG LLT) system. Using ~3 million pairs of turbulence diagnostics and in situ eddy dissipation rate observations, we trained and evaluated random forest, Extreme Gradient Boosting, and Light Gradient Boosting Machine models. All three consistently outperformed GTG LLT but shared limitations in seasonal, diurnal, and altitude-dependent performance patterns. SHapley Additive exPlanations analysis was applied to interpret diagnostic contributions, offering clues on the processes influential for turbulence prediction. To refine performance, three strategies were introduced: (i) threshold adjustment, (ii) regression-adapted Synthetic Minority Over-sampling Technique to address class imbalance in rare turbulence events, and (iii) quantile regression with tree ensembles to produce predictive intervals and quantify spatially varying uncertainty critical for safety-critical aviation operations.