Covariance Localization in Strongly Coupled Data Assimilation

Takuma Yoshida

  2 August, 1pm, in 2890


Coupled models of the Earth system have now enabled numerical prediction from weather time scales to climate projections. Strongly coupled data assimilation (DA) based on an ensemble of forecasts is a promising approach for providing initial conditions for these coupled models due to their ability to estimate flow-dependent coupled error covariance. Because the coupling strength between subsystems of the Earth is not a simple function of a distance, we need a better localization strategy than the current distance-dependent localization.

We first propose the correlation-cutoff method, where localization of strongly coupled DA is guided by ensemble correlations of an offline DA cycle, so that, for example, an atmospheric observation will be assimilated into an ocean location only if the variables at the two locations have been determined to have significantly correlated errors. The method improves the analysis accuracy when tested with a simple coupled model of atmosphere and ocean. We then extend the correlation-cutoff method to a global atmosphere-ocean strongly coupled DA with neural networks. The combination of static information provided by the neural networks and flow-dependent error covariance estimated by the ensemble improves the atmospheric analysis in our observation system simulation experiment. The neural networks can reproduce the global error statistics reasonably well, and their computational cost in a DA system is reasonable.

As a related topic, error growth and predictability of a coupled dynamical system with multiple timescales are explored with a simple coupled atmosphere-ocean model. The attractor is found to have a discontinuous response to the strength of the coupling.