Item

Abstract

Rapid advances in machine learning (ML) surrogate models of the atmosphere have allowed such models to obtain similar or greater forecast accuracy to state-of-the-art numerical models for certain prognostic variables. While these ML surrogates are often trained on long time-series of reanalysis data, recent work has shown that cyclic data assimilation may instead be used to simultaneously learn ML model parameters and estimate the model state, allowing for online training using partial and noisy observations. Machine learning has additionally been shown to be able to learn the observation operator used in the data assimilation cycle. Combining these two ideas, we propose a technique for using cyclic data assimilation, in the form of the ensemble transform Kalman filter (ETKF), to simultaneously (a) learn a hybrid ML model for the bias correction to both the observation operator and the model forecast and (b) to estimate the system state. We test this technique using a "reservoir computer" as our hybrid model's ML component and the Lorenz Model 1 as our hybrid model's knowledge-based component. We test our model on two cases: in the first, observations are obtained from the single-scale Lorenz Model 2, while in the second, observations are obtained from the coarse-scale variables of the multi-scale Lorenz Model 3.