Machine Learning-Driven Background Error Covariances for High-Resolution Data Assimilation

In high-resolution Numerical Weather Prediction (NWP) and data assimilation, producing an efficient analysis relies heavily on background error covariances. Because these errors are highly flow-dependent, capturing them traditionally requires, e.g., generating large ensembles, which is computationally challenging for urban-scale models—particularly regarding vertical error covariances. To address this, we are developing a machine learning surrogate method to approximate flow-dependent covariances in the vertical direction. Preliminary results using fully connected neural networks indicate that the model can successfully learn error structures and reproduce them using the vertical profiles of a single forecast, potentially reducing the reliance on computationally expensive ensembles.