Abstract

Political decisions, adaptation planning, and impact assessments need reliable estimates of future climate change and related uncertainties. To provide these the spread in multi-model projections (for example from CMIP6) is often translated into probabilistic estimates such as the mean and likely range. However, considering only the raw model distribution has several potential shortcomings. To address these, a model weighting scheme, which accounts for the models’ historical performance as well as model interdependence within the multi-model ensemble, is introduced. It is shown that models known to be structurally similar can be clustered in a “model family tree” based solely on their output fields. Independence weights are then derived for all models based on the degree of dependence between each model pair to correct, for example, for shared components. Model performance compared to observations is investigated based on several metrics and then translated into performance weights. Applying the combined performance-independence weights to projections of global mean temperature change from CMIP6 leads to reduced warming as well as reduction in uncertainty. Different ways to ensure the quality of the weighting compared to the unweighted case are discussed with a focus on comparing it to a range of other methods. It is shown that there is general agreement between different constraining methods, depending on a range of factors.