Summary and Recommendations#
Summary#
A 24-year (1999–2022) hindcast simulation for the Cook Inlet Operational Forecast System (CIOFS) model has been run and made publicly accessible 1 in order to meet the goal of the Cook Inlet Circulation Modeling project to develop the resources and tools needed to calculate oil spill trajectories. The Cook Inlet Circulation Modeling project contributes to the Oil Spill Recovery Institute (OSRI) main goals to understand Arctic and sub-Arctic marine environments, inform oil spill prevention and response, as well as enhance oil spill response efforts. Outputs from this effort will help drive future development of particle trajectory tools for oil spill response planning, resource management, and data visualizations to better meet the needs of marine pilots and industry in Cook Inlet.
The operational version of CIOFS 2 is run by the National Ocean Service Center for Operational Oceanographic Products and Services and generates nowcasts and 48-hour forecasts 4 times per day and generally captures the highly-variable bathymetry and tides of Cook Inlet making it potentially well-suited for coastal planning and resource management applications. However, the model currently uses limited freshwater forcing and its accuracy has been unclear for subtidal currents, temperature, and salinity patterns, all of which are important for ecological applications and oil spill response. To address these gaps, the Cook Inlet Circulation Model project, in addition to generating the hindcast 3-dimensional oceanographic model output, has assessed hindcast CIOFS skill by comparing it to the Northwest Gulf of Alaska (NWGOA) three-dimensional ocean circulation model 3 and validated the CIOFS hindcasts using physical measurements to assess the accuracy of model results.
Key Outcomes#
The model-model comparisons (CIOFS to NWGOA) show CIOFS to be saltier and less variable than NWGOA most of the time. CIOFS is colder than NWGOA in the winter and hotter in the summer. The higher horizontal resolution in CIOFS allows for improved modeling of tidal rips as compared with NWGOA.
The model-data comparisons indicate CIOFS lacks the magnitude of freshwater input into the system to have realistic salinity variability or accurate salinity. NWGOA seems to have enough freshwater input, leading to realistic variability, but not enough horizontal resolution to accurately calculate its transport. The major seasonal sea water temperature signal is well-captured in both models, but temperature anomaly is not well-captured. CIOFS is better at accurately modeling cross-channel variation than NWGOA, possibly due to higher horizontal resolution, and NWGOA is better at vertical variation in temperature, possibly due to higher vertical resolution. Both models mostly capture both the tidal and subtidal sea surface height well. Additionally, both models capture along-channel surface currents in two channels, but mostly do poorly with across-channel currents as well as subtidal in either direction. CIOFS performs better than NWGOA for horizontal speed so is capturing the kinetic energy in the model but missing the directionality which may be related to local bathymetric gradients that are not in the bathymetry data.
Recommendations#
The CIOFS model already has decent horizontal resolution, especially compared to NWGOA, but needs much more freshwater input to the system to potentially be able to accurately model the salinity field. Increasing the CIOFS model vertical resolution would improve the model’s ability to capture vertical processes.
The CIOFS hindcast model requires almost 100TB of storage space, presenting a challenge for serving it. The CIOFS hindcast model, CIOFS operational model, and NWGOA are all available through a THREDDS Data Server (TDS), a typical way to serve model output. We also experimented with serving the model output using xpublish 4, which provides a modern approach for serving files via multiple API endpoints. This approach has potential but is also still under active development. Assuming development goes well, we expect to serve through xpublish more over time.
The CIOFS hindcast simulations were performed using the Intel Fortran Compiler and Intel MPI (provided in Intel oneAPI base toolkit and HPC toolkit, version 2023) for parallelization on a local cluster consisting of 9 nodes (40 Intel Xeon Gold 6148 CPU cores), enabling the 24-year hindcast to run in approximately 7 months. This performance demonstrates efficient scaling to 360 cores; the Intel compiler and MPI implementation tailored for the Xeon architecture was critical in maximizing utilization of the available compute capability.