Data-driven stochastic Lie transport modelling of the 2D Euler equations

TL;DR

This paper develops data-driven stochastic parametrizations for the 2D Euler equations using coarse-grid SPDEs, employing EOF-based stochastic processes that improve uncertainty quantification and prediction accuracy over Gaussian noise models.

Contribution

It introduces a novel EOF-based stochastic forcing framework for the 2D Euler equations that better captures data properties and reduces uncertainty in coarse-grid models.

Findings

Reduced uncertainty in stochastic ensembles.

Improved short-term prediction accuracy.

EOF-based models outperform Gaussian noise models.

Abstract

In this paper, we propose and assess several stochastic parametrizations for data-driven modelling of the two-dimensional Euler equations using coarse-grid SPDEs. The framework of Stochastic Advection by Lie Transport (SALT) [Cotter et al., 2019] is employed to define a stochastic forcing that is decomposed in terms of a deterministic basis (empirical orthogonal functions, EOFs) multiplied by temporal traces, here regarded as stochastic processes. The EOFs are obtained from a fine-grid data set and are defined in conjunction with corresponding deterministic time series. We construct stochastic processes that mimic properties of the measured time series. In particular, the processes are defined such that the underlying probability density functions (pdfs) or the estimated correlation time of the time series are retained. These stochastic models are compared to stochastic forcing based on Gaussian noise, which does not use any information of the time series. We perform uncertainty quantification tests and compare stochastic ensembles in terms of mean and spread. Reduced uncertainty is observed for the developed models. On short timescales, such as those used for data assimilation [Cotter et al., 2020], the stochastic models show a reduced ensemble mean error and a reduced spread. Particularly, using estimated pdfs yields stochastic ensembles which rarely fail to capture the reference solution on small time scales, whereas introducing correlation into the stochastic models improves the quality of the coarse-grid predictions with respect to Gaussian noise.