A STOCHASTIC-MODEL OF SST FOR CLIMATE SIMULATION EXPERIMENTS

This study describes the implementation of a statistical method to sim ulate a multi-century sequence of global sea surface temperature (SST) fields. A multi-variable auto-regressive (AR) model is trained on the observed time series of SST from the data set compiled at the Hadley Centre (GISST 2.0). To reduce the dimensionality of the model, the sto chastic process is in practice fitted to empirical orthogonal function (EOF) time coefficients of the SST series, retaining the first 14 EOF s. Selected lag cross-covariances among the EOF time series are retain ed, based on the structure of the cross-correlation matrix and lags up to 64 months are included. Though the resulting system is quite large (a 14-dimensional AR process, with 400 parameters to be determined) t he calculation is possible and a stable process is obtained. The proce ss can then be used to investigate some statistical properties of the SST data set and to generate synthetic SST data that could be used in very long numerical experiments with atmospheric or ocean models in wh ich only the main features of the observed statistics of the SST must be retained. Results indicate that the synthetic SST data set seems to be of usable quality as boundary condition for the atmosphere or the ocean in climate experiments. Analysis of extreme events and extreme d ecades in the synthetic SST data confirms the exceptional character of the 1980s, but also provides circumstantial evidence that the 1980s w ere indeed within the limits of the statistics of the previously obser ved record.