More accurate quantification of model-to-model agreement in externally forced climatic responses over the coming century

Separating how model-to-model differences in the forced response (U-MD) and internal variability (U-IV) contribute to the uncertainty in climate projections is important, but challenging. Reducing U-MD increases confidence in projections, while U-IV characterises the range of possible futures that might occur purely by chance. Separating these uncertainties is limited in traditional multi-model ensembles because most models have only a small number of realisations; furthermore, some models are not independent. Here, we use six largely independent single model initial-condition large ensembles to separate the contributions of U-MD and U-IV in projecting 21st-century changes of temperature, precipitation, and their temporal variability under strong forcing (RCP8.5). We provide a method that produces similar results using traditional multi-model archives. While U-MD is larger than U-IV for both temperature and precipitation changes, U-IV is larger than U-MD for the changes in temporal variability of both temperature and precipitation, between 20 degrees and 80 degrees latitude in both hemispheres. Over large regions and for all variables considered here except temporal temperature variability, models agree on the sign of the forced response whereas they disagree widely on the magnitude. Our separation method can readily be extended to other climate variables. Uncertainty in estimates of future climate arises not only from internal variability, but also from model-to-model differences. Here, the authors use a new set of single model initial-condition large ensembles to quantify the contribution of model differences to the overall uncertainty in temperature and precipitation projections.

Automated summary

Understanding the contributions of model-to-model differences in the forced response and internal variability to uncertainty in climate projections is crucial. While reducing model-to-model differences can increase confidence in projections, separating these uncertainties is challenging in traditional multi-model ensembles. However, utilizing single model initial-condition large ensembles can help to distinguish between these uncertainties.