[M-GI26] Data assimilation: A fundamental approach in geosciences

[MGI26-P02] Implementation of the local particle filter as an extension of the SCALE-LETKF with weight interpolation

★Invited Papers

*Arata Amemiya^1,2, Takemasa Miyoshi^1,2,3 (1.RIKEN Center for Computational Science, 2.RIKEN Cluster for Pioneering Research, 3.RIKEN Interdisciplinary Theoretical and Mathematical Sciences Program)

Keywords:Data Assimilation, Particle filter, Numerical weather prediction

The particle filter is an ensemble data assimilation method which does not assume a Gaussian distribution. There is a growing attention to the application of the particle filter to atmospheric and ocean data assimilation, since nonlinear and non-Gaussian nature of the system often makes the Kalman filter formulation suboptimal. The application of the particle filter to large-dimensional spatiotemporal systems is generally challenging. Various methods to tackle the large dimensions have been proposed such as spatial localization and hybrid approaches with the ensemble Kalman filter. Recently, an efficient implementation of the local particle filter (LPF) with minor modifications of the local ensemble transform Kalman filter (LETKF) has been proposed by Kotsuki et al. (2022). We implemented the LPF to the regional numerical weather prediction model SCALE-RM. We also implemented a spatial interpolation of the transformation matrices to reduce the computational cost and to improve consistency between neighboring grid points. We performed preliminary experiments using idealized settings of large-scale baroclinic waves and showed that the SCALE-LPF worked stably with the analysis error comparable to that of LETKF. The SCALE-LPF showed a clear advantage over LETKF in the scaling of computational time with respect to the ensemble size.