Sequential Observation Assimilation using an Ocean 4DVAR System in Reduced Numerical Precision

The Navy’s ocean data assimilation approach for sea surface height (SSH) observations relies on deriving synthetic profiles of temperature and salinity that, when assimilated, adjust the ocean model to match the SSH observations. This process uses a method known as the Improved Synthetic Ocean Profiles (ISOP) system and its accuracy depends partially on an accurate representation of sea surface temperatures (SST). In practice, the reliability of this process is limited by the accuracy of the SSTs that are provided by the background model forecast. An alternative approach relies on the use of outer loops in the assimilation process, where the first outer loop assimilates SST observations and the resulting analysis is then used in a second outer loop as the updated background state for assimilation of ISOP profiles. A third outer loop can then be performed to assimilate any available in-situ profiles. Typically, this process can be very costly especially for a four-dimensional variational (4DVAR) method. However, the computational cost can be reduced by solving the 4DVAR using reduced numerical precision. This approach is applied to the Navy Coastal Ocean Model (NCOM) 4DVAR in a real-world application of the Northeast Atlantic domain in June-August, 2019. Experiment results show that the single-precision 4DVAR is able to perform in a nearly identical manner with double-precision 4DVAR in terms of analysis and subsequent forecast accuracy. The single-precision 4DVAR is able to produce an analysis, however, in roughly half the time of the double-precision 4DVAR. This cost savings makes the multiple outer loop approach computationally feasible. The results from the three outer loop experiment show substantial improvement in the 24-hr forecast of sub-surface temperature and salinity when compared to a run using a single outer loop where all observations are assimilated simultaneously.