Influence of Parallel Computational Uncertainty on Simulations of the Coupled General Climate Model

Parallel computational uncertainty, arising from the non-associativity of floating-point arithmetic in parallel reductions, can introduce perturbations in climate model simulations. We investigate this effect in a coupled general circulation model. Results show that changing the parallel domain decomposition or processor count induces round-off errors that grow rapidly, leading to significant divergence in the simulated climate state after a short period. This uncertainty is comparable to that caused by initial condition perturbations. The study emphasizes the need to account for computational uncertainty when interpreting model results and performing ensemble simulations.

Geoscientific Model Development