Machine Learning for Turbulence Risk Identification in U.S. Airspace Using Open Flight and Weather Data

Turbulence is a persistent and often unpredictable hazard in aviation, frequently occurring without visual or radar cues.This study presents a machine learning framework for identifying severe and extreme turbulence risk across U.S. airspace using only publicly available flight and weather data.The framework combines over 550,000 pilot reports with ERA5 reanalysis data to construct a large labeled dataset.It integrates anomaly-aware downsampling, synthetic oversampling, dimensionality reduction, and both unsupervised (K-Means) and supervised (XGBoost) modeling.In 10-fold cross-validation, the model achieved strong performance (recall = 0.91, F1 = 0.88) in detecting high-risk events.A real-world case study from February 2025 further illustrates the system's predictive capability.This work demonstrates the feasibility of operational turbulence identification using open-source data and interpretable learning techniques.