Paper detail

Using reduced-precision arithmetic in the adjoint model of MITgcm

In recent years, it has been convincingly shown that weather forecasting models can be run in single-precision arithmetic. Several models or components thereof have been tested with even lower precision than this. This previous work has largely focused on the main nonlinear `forward' model. A nonlinear model (in weather forecasting or otherwise) can have corresponding tangent linear and adjoint models, which are used in 4D variational data assimilation. The linearised models are plausibly far more sensitive to reductions in numerical precision since unbounded error growth can occur with no possibility of nonlinear saturation. In this paper, we present a geophysical experiment that makes use of an adjoint model to calculate sensitivities and perform optimisation. Using software emulation, we investigate the effect of degrading the numerical precision of the adjoint model. We find that reasonable results are obtained with as few as 10 significand bits, equal to the significand precision in the IEEE half-precision standard.