Algorithmic Differentiation for Climate Science
|Title||Algorithmic Differentiation for Climate Science|
|Year of Publication||2016|
|Authors||Narayanan, SHK, Goldberg, N, Heimbach, P, Marshall, D, Price, S, Wunsch, C|
Computing sensitivities of climate-related quantities of interest (QoI) to very high-dimensional state or parameter spaces in an efficient manner is of considerable interest in climate science. Applications in which sensitivities are propagated in time by means of adjoint models include quantifying controlling factors of the QoI’s variations [24, 3], optimal state and parameter es- timation (”data assimilation”) [25, 26], computing nonnormal transient amplification patterns , and formal uncertainty quantification [10, 11]. In an environment of rapid model de- velopment algorithmic differentiation (AD) plays a crucial role in supporting adjoint model generation for diverse applications based on up-to-date source codes . With applications ranging from oceanography , atmospheric sciences , coupled carbon cycle modeling  and cryospheric sciences [5, 2], the need for sustained AD tool maintenance and improvements becomes a high priority.