Advancements in numerical weather prediction (NWP) using machine learning

The introduction of artificial intelligence and machine learning (AI/ML) methods has completely changed the nature of Numerical Weather Prediction (NWP). This paper is a systematic analysis of the level of AI-enhanced model excellence in forecasting skills as compared to conventional physics-based NWP models especially on extreme weather events, such as hurricanes and heatwaves. The traditional NWP models (GFS and ECMWF) and AI-based (CNN-LSTM) as well as hybrid NWP-AI models are evaluated using the comparative empirical evaluation of the results using historical and real-time meteorological data. Findings have shown that AI-based models experience significant and statistically significant forecast error reduction with an average absolute error and root mean square error reducing by up to 30-40% compared to conventional NWP forecasts. The overall performance of hybrid NWP-AI models is the most effective, with the lowest error variance and greatest detection ability of events, which is demonstrated by F1-scores above 0.85. The error decomposition and residual variance were used to further exclude that the AI integration decreases the forecast uncertainty substantially, especially on long lead times where the classical models have high error growth rates. Hurricane track prediction and heatwave-intensity prediction case studies expose that AI-boosted forecasting is more effective in the prediction of storm routes, as well as storm peaks and temporal development of extremes, and are more consistent with observed data. Besides improved accuracy, AI-based and hybrid methods have shown significant improvements in computational efficiency that can be used to make inferences quicker and avoid energy-consuming high-resolution simulations. The results prove that AI-enhanced and hybrid NWP systems can provide simultaneous gains in accuracy, reduction of uncertainty, and computational efficiency, and can be used to improve operational weather forecasting, and increase resilience to climate-driven extreme events. Future research will focus on Physics-informed machine learning, uncertainty quantification, and real-time data assimilation frameworks.