Research on lightweight machine learning for aviation weather forecasting across 11 international airports Key results: → 2.5-4x improvement vs operational forecasts → AUC 0.89-0.98 across diverse climates → Runs on commodity hardware (~900KB models) → SHAP reveals learned physics (advection, radiation, subsidence) Using only publicly available data, the framework enables any airport to deploy advanced forecasting in minutes—no proprietary data, no expensive infrastructure.