Ensemble filtering with displacement errors

Michael Ying
NCAR, Boulder, Colorado
Seminar Date: 
12. February 2020 - 13:00 - 14:00
Lecture room, Ground Floor, NERSC

An outstanding issue for multiscale weather prediction is the choice of data assimilation methods. Since small scale features rapid error growth that gives rise to nonlinearity, data assimilation methods based on linearization, such as the ensemble Kalman filter (EnKF), performs suboptimally. Position error of convective clouds among ensemble members is one of the common causes of nonlinearity and has been a major challenge for data assimilation. Previous studies have made some progress in developing nonlinear optimization methods to reduce position errors. However, these methods are costly and not yet practical for large dimensional problems. I developed a multiscale alignment method that decompose the model state into scale components, and solve the data assimilation problem iteratively from large to small scales. The new algorithm utilizes large-scale analysis increment to compute a displacement vector field, which removes the position errors at smaller scales. The new multiscale algorithm has been tested as a proof of concept using a quasi-geostrophic model, and is currently being implemented in Data Assimilation Research Testbed (DART) for future application in more complex systems.