Multi-Decadal Prediction Stream
An important question in the study of climate change is whether predictions of future climate change could be improved if models start their predictions from an observed state of the climate system. The underlying issue is that changes in the climate system are a combination of internal variability of the coupled system and the response of the climate system to radiative forcing changes (the forced response). Most previous climate change simulations have started from an arbitrary point in a long control simulation, and then impose changing atmospheric composition. The detailed time evolution of the model response is not expected to match the observed seasonal to decadal evolution of the climate system, since the model’s internal variability will not match that of the real climate system. This type of simulation is only meant to estimate the forced response of the climate system.
As part of the CMIP5/IPCC AR5 assessment, the international community is conducting a set of coordinated experiments in which models used for the prediction of climate change are initialized with estimates of the observed state of the climate system. The key question is whether this initialization process produces model simulations and predictions that are more skillful at predicting the details of the future evolution of the climate system across time scales, from seasonal to decadal and longer. These simulations are designed to compute both the forced response of the climate system and the time evolution of the internal variability of the climate system from an observed state.
GFDL is actively participating in this internationally coordinated activity. We have used the GFDL CM2.1 climate model (Delworth et al., 2006) to conduct an extensive set of hindcasts and predictions. The model is initialized using a state of the art coupled assimilation system (Zhang et al., 2007). For each set of observed states (initial conditions) we conduct a ten member ensemble of ten-year hindcasts or predictions (hindcasts when the starting date is in the past, and so its accuracy can be assessed; predictions when the starting date is close to the present, and so its accuracy is not yet known).
We use observed states from January 1 of each year from 1960 to 2011, for a total of 52 hindcasts and predictions (representing 5200 model simulated years). The ensemble members differ slightly in their initial conditions as derived from the assimilation system. The simulations use estimates of observed changes in radiative forcing until 2005, and estimated forcing according to the RCP4.5 scenario thereafter. We refer to these simulations as the initialized experiments.
In addition, we have conducted a separate 10-member ensemble of simulations using the CM2.1 model that are not initialized from an observed state. These start from arbitrary initial conditions in a control simulation, and cover the period 1861-2040. The simulations use the same changes in radiative forcing that the hindcasts use, and are intended to estimate the forced response of the climate system. We refer to these as the uninitialized experiments. The central question to ask is whether the initialized experiments provide better hindcasts than the forced experiments.
Delworth, T.L., et al., 2006: GFDL’s CM2 Global Coupled Climate Models. Part I: Formulation and Simulation Characteristics. Journal of Climate, 19(5), DOI: 10.1175/JCLI3629.1.
Zhang, S., M.J. Harrison. A. Rosati, and A.T. Wittenberg, 2007: System Design and Evaluation of Coupled Ensemble Data Assimilation for Global Oceanic Climate Studies. Monthly Weather Review, 135(10), DOI:10.1175/MWR3466.1.