Abstract:

The adjoint model of the GRAPES regional model has been developed with a combination of automatic differentiation tool TAPENADE and handcoding. The model is made up of mode programs and a memory supporting library (PLP library). The TAPENADE employs a complete storage strategy to save the intermediate variables. This causes too much memory overhead during the operation of the adjoint mode and increases memory consumption with the time step going on. The growing of memory is addressed by modifying the POP functions in the PLP. The mode programs are optimized for integer variables, real variables and calling subroutines of the forward model so that the memory overhead is significantly decreased.

Key words: numerical weather forecast, GRAPES, adjoint model;memory optimization;TAPENADE