Machine learning-based estimation of state-dependent forecast uncertainty

JUAN RUIZ; MAXIMILIANO SACCO; MANUEL PULIDO; PIERRE TANDEO

Tokyo

Congreso; 10th International Congress on Industrial and Applied Mathematics; 2023

Quantifying forecast uncertainty is a key aspect of state-of-the-art numerical weather prediction and data assimilation systems. State dependent uncertainty quantification in numerical weather prediction is a computation intensive task which has been performed using different approaches such as monte carlo sampling (ej. ensemble Kalman filter) and variational approaches (ej. adjoint based model sensitivity). Machine learning techniques consist of trainable statistical models that can represent complex functional dependencies among different groups of variables given a large enough dataset. In this talk we will describe the use of a machine learning approach based on neural networks for the estimation of forecast uncertainty. In particular, we will discuss the estimation of the forecast error covariance matrix, which is at the center of probabilistic forecasting and data assimilation systems. In addition, we will present a hybrid data assimilation method that combines the optimal interpolation technique and a convolutional neural network to estimate the state dependent forecast error covariance matrix.