9:30 AM - 9:45 AM
[AAS07-03] 1000-member 18-km-mesh SCALE-LETKF experiment with conventional observations in summer 2020
Keywords:Data Assimilation, Numerical weather forecasting, Big data
Advances of high-performance computing technologies have enabled us to study ensemble simulations with unprecedently large ensemble sizes. For example, a recent study performed a 1000-member convective-scale ensemble forecast experiment to examine the forecast error properties (Necker et al. 2020). Ensemble-based data assimilation such as the local ensemble transform Kalman filter (LETKF; Hunt et al. 2007) with a large ensemble size is also an important topic to study. The K computer played an important role in early works of such large ensemble data assimilation (Miyoshi et al. 2014; 2015, Kondo and Miyoshi 2016). More recently, 1024-member data assimilation with the 3.5-km resolution global model (Nonhydrostatic Icosahedral Atmospheric Model; NICAM) was performed on the supercomputer Fugaku (Yashiro et al. 2020). The previous studies demonstrated the advantage of a large ensemble in idealized experiments, but it is still unclear if the advantage is also true in the real situation, where imperfections are included in the observation and model, and in representing the observation errors. This study examines the advantage of a large ensemble of 1000 members in the LETKF in a real case with conventional observations in the Japan region in summer 2020.
We used the SCALE-LETKF system (Lien et al. 2017) composed of the 18-km resolution Scalable Computing for Advanced Library and Environment-Regional Model (SCALE-RM, Nishizawa et al. 2015) and the LETKF. The conventional observation dataset known as PREPBUFR used in NCEP Global Data Assimilation System (GDAS) is assimilated every 6 hours. For the lateral and surface boundary conditions, NCEP Global Forecasting System (GFS) data is used. In this study, we run the SCALE-LETKF with ensemble sizes up to 1000. The covariance inflation and localization parameters are fixed regardless of the ensemble size, so that we focus on the qualitative impact on the analysis. For each experiment with a different ensemble size, the model is initialized at 0000 UTC 1 August 2020, with NCEP GDAS field superposed with bandpass-filtered random temperature perturbations. The analyses at 0000 UTC 3 August are investigated.
The spatial and cross-variable correlation signals become clearer as the ensemble size increases (Figure 1). The analysis RMSE relative to the observations is significantly reduced for most variables and vertical levels with larger ensemble sizes. The first-guess RMSE is reduced as well, implying that the analysis field of the previous step is improved in a physically consistent manner.
However, we observed some undesirable behaviors as well. The analysis RMSE relative to surface pressure observations increased with larger ensemble sizes, suggesting that the LETKF fail to alleviate the surface pressure bias. An additional experiment with 1000-member LETKF without vertical localization did not show improvement of the analysis for most variables, on the contrary to the theoretical expectation. The possible cause of this behavior is the systematic model and observation biases. The treatment of those biases would be important to achieve the full potential of the LETKF with a large ensemble size. The results of further experiments will be discussed in the presentation.
Acknowledgement: The computation for this study was performed using Oakforest-PACS through HPCI System Research Project (Project ID: hp150019).
We used the SCALE-LETKF system (Lien et al. 2017) composed of the 18-km resolution Scalable Computing for Advanced Library and Environment-Regional Model (SCALE-RM, Nishizawa et al. 2015) and the LETKF. The conventional observation dataset known as PREPBUFR used in NCEP Global Data Assimilation System (GDAS) is assimilated every 6 hours. For the lateral and surface boundary conditions, NCEP Global Forecasting System (GFS) data is used. In this study, we run the SCALE-LETKF with ensemble sizes up to 1000. The covariance inflation and localization parameters are fixed regardless of the ensemble size, so that we focus on the qualitative impact on the analysis. For each experiment with a different ensemble size, the model is initialized at 0000 UTC 1 August 2020, with NCEP GDAS field superposed with bandpass-filtered random temperature perturbations. The analyses at 0000 UTC 3 August are investigated.
The spatial and cross-variable correlation signals become clearer as the ensemble size increases (Figure 1). The analysis RMSE relative to the observations is significantly reduced for most variables and vertical levels with larger ensemble sizes. The first-guess RMSE is reduced as well, implying that the analysis field of the previous step is improved in a physically consistent manner.
However, we observed some undesirable behaviors as well. The analysis RMSE relative to surface pressure observations increased with larger ensemble sizes, suggesting that the LETKF fail to alleviate the surface pressure bias. An additional experiment with 1000-member LETKF without vertical localization did not show improvement of the analysis for most variables, on the contrary to the theoretical expectation. The possible cause of this behavior is the systematic model and observation biases. The treatment of those biases would be important to achieve the full potential of the LETKF with a large ensemble size. The results of further experiments will be discussed in the presentation.
Acknowledgement: The computation for this study was performed using Oakforest-PACS through HPCI System Research Project (Project ID: hp150019).