[M-GI24] Data assimilation: A fundamental approach in geosciences

[MGI24-04] Quantifying the relationships between parameter identifiability and parameter ensemble spread in the DA-based parameter estimation: An ideal 2D squall-line experiment

*江 嘉敏¹、雨宮 新¹、山浦 剛¹、末木 健太²、富田 浩文¹ (1.理化学研究所、2.気象庁気象研究所)

キーワード：Data assimilation、Parameter estimation、Parameter identifiability、Parameter ensemble spread 、Microphysical schemes、An ideal 2D squall-line experiment

Parameter estimation (PE) using the ensemble-based data assimilation technique has been developed and practiced in the physical schemes for weather-climate simulations (e.g., Kotsuki et al., 2018, 2020; Sueki et al., 2022). However, almost all these parameters cannot be directly measured, and they are significantly correlated with the observation operator and the quality of the observational data. Although previous studies found that the performance of single and multiple PE(s) are strongly related to the parameter identifiability based on the observation types, a suitable parameter ensemble spread (ES) also plays an important role in the estimation efficiency, such as a higher precision and faster convergence speed of the estimation. Therefore, in this study, we examined and quantified the relationships between parameter identifiability and estimation performance with different constant parameter ensemble spreads. Here, we implemented the parameter estimation algorithm into a regional atmospheric DA system (SCALE-LETKF), which incorporates the Scalable Computing for Advanced Library and Environment-Regional Model (SCALE-RM) and the Local Ensemble Transform Kalman Filter (LETKF). A two-dimensional ideal squall-line observing system simulation experiment (OSSE), in which synthetic radar reflectivity is assimilated using the observation operator of Amemiya et al. (2020), was examined. Multiple parameters in microphysical schemes, that is empirical coefficients of rainfall and snow and drag coefficient of graupel (C_r, C_s, and drag_g) in Tomita08, were simultaneously estimated under a nondimensional constant parameter ensemble spread in the tanh-transformation space. In the results of single PE, we found that C_r is the most identifiable, which means the fastest convergence speed and the highest precision at the same ES condition, when C_s is weakly correlated with the reflectivity. a better convergence speed is demonstrated if ES is 0.05—0.10 when the precision becomes low and stable with ES < 0.1 in the cases of C_r and drag_g. In the case of C_s, smaller ES has better precision, and a faster convergence speed despite a relatively low precision if ES=0.08—0.10. Multiple (three) PEs showed that the performance of the high-identifiable parameter (C_r) is similar to its single PE while the lower precision and slower convergence speed are presented in the other two parameters (C_s, and drag_g), compared with their single PEs. Moreover, the difference in convergence speed between single and multiple PEs becomes large with a small parameter ES.