Tangent Linear and Adjoint Models of the Cubed-Sphere GEOS GCM

The Global Modeling and Assimilation Office (GMAO) Goddard Earth Observing System (GEOS) data assimilation system (DAS) uses an adjoint (AD) general circulation model (GCM) as part of a procedure to calculate the impact of observations used in the DAS on 24-hour forecasts. This AD model and its correspondent tangent linear (TL) model have been developed based on an early version of the fully nonlinear GEOS-GCM that uses a one-dimensional MPI domain-decomposed finite-volume hydrodynamics, and has thus corresponding limited scalability. In order to improve upon parallel scalability, the TL and AD models have been re-derived using the more recent, cubed-sphere, GEOS GCM hydrodynamics which, in particular, follows a more effective two-dimensional MPI domain decomposition.

The initial derivation employs the Transformation of Algorithms in FORTRAN (TAF) automatic adjoint generation tool. Basic features such as connections to the Earth System Modeling Framework and to GFDL's Flexible Modeling System have been hand-treated to maintain parallel scalability and consistency with the MPI communications to those of the nonlinear hydrodynamics. Additionally, the interplay between recomputation and checkpointing, largely affecting memory usage, is being addressed in different ways, namely by: investigating various options of the piece-wise parabolic method participating in the dynamics; implementing simplifications to the internal time-split option; and pre-calculating interpolation coefficients involved in the transformations converting points on the cubed-grid to their closest latitude-longitude on a regular grid (these being required to interface the TL and AD models with the Grid-point Statistical analysis system used in the GEOS DAS). Along with various standard TL and AD sanity checks, results using the cubed-sphere AD model in the observation impact calculations will be discussed in this presentation.