Assessing a Technique for Integrated Data Assimilation and Machine Learning with a Hybrid Atmospheric Model

Even a state-of-the-art numerical weather prediction (NWP) model is an imperfect model of the atmosphere. A highly promising recently introduced approach to reduce forecast errors associated with imperfections of an NWP model is to augment the physics-based NWP model with a machine learning (ML) based component. The resulting hybrid model uses observational analyses both for the training of the ML component and as the initial conditions of forecasts. Our long-term goal is to develop a methodology in which the training of the ML component and the data assimilation that provides the observational analyses are fully integrated. In this talk, we asses the performance of a simple approach, which is our first step of developing such a methodology. In this approach, a long time series of analyses is produced first, assimilating past observations of the atmospheric state, using the NWP model to produce the 6-hour forecasts that provide the background states. The hybrid model is then trained on this time series of analyses. Next, a new time series of analyses is prepared, but this time using the trained hybrid model to provide the background states. The last two steps can be repeated until the analyses cannot be further improved by retraining the hybrid model. The performance of this approach is assessed by experiments in which the low-resolution atmospheric general circulation model SPEEDY serves as the NWP model, the approach of Aromano et al. (2022) is used to hybridize the model, the observations are pseudo-observations based on ERA5 reanalyses, and the Local Ensemble Transform Kalman Filter (LETKF) is used for data assimilation. It is shown that this iterative approach produces overall more accurate analyses than the conventional approach in which the NWP model is used to produce the background. The improvements are the largest for the state variables that are most affected by the errors of the NWP model.