Estimating JEDI Based MOM6 3DVAR Covariance for HAFS

Estimating JEDI Based MOM6 3DVAR Covariance for HAFS

Yongzuo Li, HeeSook Kang, Kriti Bhargava, Biju Thomas, Maria Aristizabal, John Steffen, Bin Li, Bin Liu, Travis Sluka, Hyun-Sook Kim, Zhan Zhang, Avichal Mehra

The Hurricane Analysis and Forecast (HAFS) system is two-way coupled with Modular Ocean Model v6 (MOM6). Recently, Joint Effort for Data Assimilation Integration (JEDI)-based MOM6 3DVAR has been developed in HAFS. This study is going to introduce the JEDI-Based MOM6 3DVAR covariance for HAFS. The 3DVAR covariance is composed of 3-dimensional spatial correlations represented by Background error on Unstructured Mesh Package (BUMP) Normalized Interpolated Convolution from an Adaptive Subgrid (NICAS) with a given length scale at horizontal and vertical directions respectively, standard deviation (StdDev, square root of the variance) interpolated from the global MOM6 model 40-member forecast ensemble StdDev, and variable balances.

Now the 3DVAR covariance is being estimated from a training data set of 300 differences (perturbations) between 48-hour and 24-hour HAFS MOM6, real time (2024) and retrospective (2020-2023), forecasts valid at the same time (00Z). The new StdDev will reflect HAFS MOM6 uncertainty more accurately. StdDev from small HAFS MOM6 samples shows reasonable patterns though much noise is seen that will be reduced by more samples. Thus better increment values are expected. The new variable balance (cross-variable linear regression coefficients) will reduce existing unreal large increments of a variable caused by observation of another variable. The horizontal and vertical correlation will be 3-dimensionally adaptive to ocean dynamics and observation distributions. Specifically, the vertical length scale is consistent with general mixed layer depth. The horizontal increment distribution patterns will match the observation patterns at sea surface as seen from our 150 km (JEDI 300 km radius setting) length scale, while the increments can be smooth between neighboring in-situ profiles at sublevels as seen from our 300 km (JEDI 600 km radius setting) length scale. All BUMP VAR, VBAL, HDIAG drivers are available for covariance generation. (example application, the Model for Prediction Across Scales – Atmosphere, MPAS).