A Mixed-Precision Hybrid Saddle-Point 4D-Var System for ROMS with Application to Assimilation of Remotely-Sensed Bio-Optical Properties

PI: Wilkin, John (Rutgers, The State University of New Jersey)
Co-PI(s): Arango, Hernan (Rutgers) : Moore, Andrew (UC Santa Cruz) : Edwards, Christopher (UC Santa Cruz) : Kurapov, Alexander (NOAA/NOS/OCS/CSDL)
Start Year: 2019 | Duration: 3 years
Partners: Rutgers, UC Santa Cruz, NOAA/NOS/OCS/CSDL

Project Abstract:

Numerical models are used in oceanography to simulate oceanic conditions for studies of regional ocean dynamics, biogeochemistry, geomorphology, and ecosystems. When operated as real-time nowcast and forecast systems, these models offer predictions that provide actionable guidance in decision-making related to water quality, public health, coastal flooding, shipping, maritime safety, and naval operations.

As is well known in the case of atmospheric modeling for weather prediction, the assimilation of observations into these models is a vital step in constraining the accuracy of the nowcast environmental conditions and thereby improving the skill of subsequent forecasts.

In recent years, tremendous progress has been made in algorithms to perform the merger of observations with hydrodynamic models that describe the time evolution of ocean conditions based on our knowledge of ocean physics. A widely used algorithm for this step, termed
“variational data assimilation” in reference to its mathematical underpinning in variational calculus, is an extremely computationally intensive task, especially if the forecast system is to achieve the high spatial resolution demanded for many naval operations and other applications, or utilize very dense data sets from emerging observing platforms.

The proposed work will significantly expand the capabilities for variational data assimilation in ROMS, the Regional Ocean Modeling System, an open-source ocean circulation model in widespread use globally for operational ocean prediction.

The new capabilities are two-fold: a significant speed-up of the underlying variational data assimilation method, and the utilization of the extensive but largely untapped (at least, operationally) data set provided by satellite ocean color imagers.

Algorithmic developments (mixed precision computing and a new “saddle-point” method for the iterative minimization of the model-data misfit cost function) will enable efficient simultaneous computation on many more computer cores of high-performance computing systems than was previously feasible, accomplishing the run-time speed-up required to address large or highresolution forecast domains. The assimilation of remote sensing reflectance in multiple wavelength bands, which are the fundamental observations returned by satellite ocean color sensors such as VIIRS, MODIS and OLCI, will be accomplished by introducing an ocean ecosystem model linked to a simple radiative transfer model that describes how light is reflected and scattered by pigments and particles in the water.

ROMS is already capable of assimilating any observations of ocean physics (salinity, temperature, velocity and sea level) that might be delivered by new sources such as satellite swath altimetry, surface current radars, or novel autonomous in situ platforms. Expanding
ROMS data assimilation capabilities to include ocean color data, and accelerating the execution speed of the data assimilation procedure, will enable the creation of data assimilative ROMS forecast systems that utilize the full suite of Global Ocean Observing System data for operational ocean prediction.

The project will be accomplished through collaboration between two academic groups (Rutgers, The State University of New Jersey, and University of California Santa Cruz), European and Australian colleagues, and the NOAA coastal marine modeling group as the federal partner who will test these advances in data assimilation capabilities in the NOAA operational environment.

BAA: N00014-18-S-B007
BAA Topic: Topic 6