Numerical Weather Prediction meets Machine Learning - a synergy for better forecasts

Numerical Weather Prediction (NWP) has recently lost its hegemony in weather forecasting, as more machine-learning-based models achieve results comparable to NWP. It turns out that data-driven models are capable of identifying patterns and distilling physical laws that, until now, have only been formulated by atmospheric physics specialists. Although ML-based models are already being used by meteorological institutes alongside NWP-based models, this does not mean that NWP will fade into irrelevance. In this talk, we will show how NWP and ML can interoperate to achieve the shared goal of providing more accurate weather forecasts. From NWP providing high-quality training data for ML models to ML models replacing specific parameterizations, the spectrum of collaboration is vast. ML models cannot succeed without high-quality training data provided by NWP, and NWP can benefit from this new technology by incorporating ML models in places where conventional physics parameterizations are found lacking. None of the currently successful ML-based models would exist without the high-quality reanalysis data generated through numerical models. Decades of expertise and extensive research in numerical modelling now serve as a solid foundation for the remarkable achievements of data-driven applications. The future of weather forecasting is built on this synergy — numerical modelling and machine learning working together to achieve what neither could accomplish on its own.