Strategies for reducing the climate noise in model simulations: ensemble runs versus a long continuous run

Abstract

Climate modelers often integrate the model with constant forcing over a long time period, and make an average over the period in order to reduce climate noise. If the time series is persistent, as opposed to rapidly varying, such an average does not reduce noise efficiently. In this case, ensemble runs, which ideally represent independent runs, can reduce noise more efficiently. We quantify the noise reduction gain by using ensemble runs over a long continuous run in constant‐forcing simulations. We find that in terms of the amplitude of the noise, a continuous simulation of 30 years may be equivalent to as few as five 3-year long ensemble runs in a slab ocean–atmosphere coupled model and as few as two 3-year long ensemble runs in a fully coupled model. The outperformance of ensemble runs over a continuous run is strictly a function of the persistence of the time series. We find that persistence depends on model, location and variable, and that persistence in surface air temperature has robust spatial structures in coupled models. We demonstrate that lag-1 year autocorrelation represents persistence fairly well, but the use of lag-1 year–lag-5 years autocorrelations represents the persistence far more sufficiently. Furthermore, there is more persistence in coupled model output than in the output of a first-order autoregressive model with the same lag-1 autocorrelation.