Augmenting covariance estimation for ensemble-based data assimilation in multiple-query scenarios

We present and assess a method to reduce the computational cost of performing ensemble-based data assimilation (DA) for reacting flows in multiple-query scenarios, i.e. scenarios where multiple simulations are performed on systems with similar underlying dynamics. The accuracy of the DA, which depends on the accuracy of the sample covariance, improves with the ensemble size, but results in a commensurate increase to computational cost. To reduce the ensemble size while maintaining accurate covariance, we propose a data-driven approach to augment the covariance based on the statistical behaviour learned from previous model evaluations. We assess our augmentation method using one-dimensional model problems and a two-dimensional synthetic reacting flow problem. We show in all these cases that ensemble size, and thus computational cost, may be reduced by a factor of three to four while maintaining accuracy.