Hybrid Weather Prediction: A Blend of Machine Learning and Numerical Modeling

Abstract:

Machine learning (ML) has been shown to skillfully predict the 3-dimensional state of the atmosphere at coarse resolution out to about 3 days. Physics-based numerical models provide skillful forecasts of the atmospheric state out to 7-10 days. We evaluate the forecast performance of a hybrid modeling approach based on combining a physics-based numerical model and a reservoir computing (RC) based ML model. We carry out forecast experiments with the Simplified Parameterization, primitivE-Equation Dynamics (SPEEDY) model of the International Centre for Theoretical Physics (ICTP) to demonstrate the approach. In these experiments, ERA-5 reanalysis data are used for the training of the hybrid model and the verification of the forecasts. The hybrid model performs better than both the numerical and the ML model for 3-6 days depending on the forecast variable. A notable feature of the hybrid forecasts is that their biases are significantly smaller than that of the numerical model.