A global perspective on CMIP5 climate model biases

The Intergovernmental Panel on Climate Change’s Fifth Assessment Report largely depends on simulations, predictions and projections by climate models1. Most models, however, have deficiencies and biases that raise large uncertainties in their products. Over the past several decades, a tremendous e ort has been made to improve model performance in the simulation of special regions and aspects of the climate system2–4. Herewe show that biases or errors in special regions can be linkedwith others at far away locations.Wefind in 22 climate models that regional sea surface temperature (SST) biases are commonly linked with the Atlantic meridional overturning circulation (AMOC), which is characterized by the northward flow in the upper ocean and returning southward flow in the deep ocean. A simulated weak AMOC is associated with cold biases in the entire Northern Hemisphere with an atmospheric pattern that resembles the Northern Hemisphere annular mode. The AMOC weakening is also associated with a strengthening of Antarctic BottomWater formation and warm SST biases in the Southern Ocean. It is also shown that cold biases in the tropical North Atlantic and West African/Indian monsoon regions during the warm season in the Northern Hemisphere have interhemispheric linkswithwarmSST biases in the tropical southeastern Pacific and Atlantic, respectively. The results suggest that improving the simulation of regional processes may not su ce for overall better model performance, as the e ects of remote biases may override them. The United Nations Intergovernmental Panel on Climate Change’s Fifth Assessment Report updates the knowledge and understanding of the scientific, technical and socio-economic aspects of climate change. The report relies heavily on the products of climate models. These, however, have serious systematic errors that challenge the reliability of climate predictions. Hence, climate model bias identification and reduction are topics of great importance. One major reason for such biases is the misrepresentations of physical processes, which can be amplified by feedbacks among climate components especially in the tropics. Much effort, therefore, is dedicated to the better representation of physical processes in coordination with intense process studies5. This paper focuses on the SST simulations by 22 participants in the Coupled Model Intercomparison Project phase 5 (CMIP5; Supplementary Information). We target the global connections among regional SST biases. The existence of such connections means that efforts to improve model performance cannot be narrowly focused on particular regions. SSTs simulated by CMIP5 models generally show too low values in the Northern Hemisphere and too high values in the Southern Hemisphere. Annual-mean SST error (that is, mean SST bias for the period from 1900 to 2005) magnitudes can be several degrees Celsius (Fig. 1a). SSTs are clearly too high in the tropical southeastern Pacific and Atlantic and too low in the equatorial and tropical southwestern Pacific. In general, these biases have patterns that are largely independent of season, but amplitudes can vary with season (Supplementary Fig. 1). For example, the warm SST bias in the Southern Ocean is present throughout the year but is much stronger during the austral summer and autumn. It is noted that the SST biases in these models are quite stable during the 1900–2005 period and the models do not show a significant SST bias trend. The misrepresentation of local processes and/or ocean– atmosphere interactions has caused some of the biases. The too warm SSTs in the tropical southeastern Pacific and Atlantic, for example, have been linked to excessive heat flux into the ocean under insufficient coverage by stratocumulus clouds6,7 combined with insufficient cooling by ocean transients from the upwelling regions along the eastern coasts8. The cold SST bias in the equatorial and tropical southwestern Pacific has been associated with an excessive westward extension of the cold tongue from the eastern equatorial Pacific in association with difficulties in the representation of surface winds and ocean mixing processes6,9. A recent study has argued that model biases even far away from the tropics can be linked to those in the tropics10. According to the study, cloud errors over the Southern Ocean may be responsible for the generation of a spurious intertropical convergence zone south of the Equator in most CMIP5 models. We start by investigating the relationships in the global domain between biases in simulated SST and in other features of atmosphere and ocean circulations. For this we take the mean AMOC as reference. The AMOC, which is characterized by warmer and saltier water flowing northward in the upper Atlantic Ocean and by cooler and fresher water flowing southward in the deep ocean11,12, is crucial to the northward heat transport by the ocean circulation13–16. As the first step in our analysis, we perform an inter-model singular value decomposition (SVD) analysis of the SST biases and AMOC streamfunction. The spatial pattern of the first SVDmode of the SST biases in Fig. 1b closely resembles the mean model biases in Fig. 1a. The corresponding AMOC mode is weakened, as indicated by the negative values of the AMOC streamfunction in the upper 3,000 m (Fig. 1c). The time series of the first SVD coefficients are highly and positively correlated (correlation coefficient 0.70). Global SST biases, therefore, strengthen as the AMOC circulation weakens. We next turn to the SST biases in the North Atlantic and Pacific oceans. It has been shown that a weakening of the AMOC is accompanied by a cooling of the North Atlantic Ocean, whereas a