Automatic Generation of Efficient Adjoint Code for the Parallel MIT General Circulation Model

In one of the most complex Earth science inverse modeling initiatives, the Estimation of the Circulation and Climate of the Ocean (ECCO) project is developing greatly improved estimates of the three-dimensional, time-evolving state of the global oceans [1, 2]. To this end, the project is applying extimation/optimal control techniques to constrain a state-of-the-art parallel general circulation model (MITgcm) [3, 4] with a diverse mix of observations, ranging from in-situ data to satellite measurements. A central ingredient of this effort is the automated generation of scalable, adjoint codes. Background Combining, through a formal optimization procedure, the fragmentary observations with a numerical model, produces an ocean state that is spatially and temporally complete and simultaneously consistent with observations and model dynamics. The difference between a model and observation is expressed in terms of a scalar cost,