Evaluation of a Hybrid Modeling Approach to Predict the Atmospheric State by Blending Numerical Modeling and Machine Learning

We evaluate the forecast performance of a hybrid modeling approach based on combining a numerical and a machine learning (ML) model. The numerical model is the Simplified Parameterizations, primitivE-Equation Dynamics (SPEEDY) model of the International Centre for Theoretical Physics (ICTP). While the dimension of the state vector of SPEEDY (~10^3) is six orders of magnitude smaller than that of a state-of-the-art numerical weather prediction model, it produces a realistic atmospheric circulation and can provide predictions of the large-scale atmospheric flow. The ML model is based on a reservoir computing approach, in which a set of reservoirs of moderate size (~10^3) predicts the atmospheric state. To be precise, each reservoir predicts the atmospheric state in a local geographical region and the calculations for the different local regions are carried out in parallel. ERA-Interim reanalysis data are used for the training of the ML model and the verification of the forecasts. It is shown, that while the ML model alone cannot predict the atmosphere, the hybrid approach provides a skillful prediction of the atmosphere for a longer time than the numerical model.