What aspect of model performance is the most relevant to skillful future projection on regional scale?

Weighting models according to their performance has been used to produce multi-model climate change projections. But the added value of model weighting for future projection is not always examined. Here we apply an imperfect model framework to evaluate the added value of model weighting in projecting summer temperature changes over China. Members of large ensemble simulations by three climate models of different climate sensitivities are used as pseudo-observations for the past and the future. Performance of the models participating in the 6th phase of the coupled model intercomparison project (CMIP6) are evaluated against the pseudo-observations based on simulating historical climatology, trends in global, regional and local temperatures, to determine the model weights for future projection. The weighted projections are then compared with the pseudo-observations in the future period. We find that regional trend as a metric of model performance yields generally better skill for future projection, while past climatology as performance metric does not lead to a significant improvement to projection. Trend at the grid-box scale is also not a good performance indicator as small-scale trend is highly uncertain. For the model weighting to be effective, the metric for evaluating the model’s performance must be relatable to future changes, with response signal separable from internal variability. Projected summer warming based on model weighting is similar to that of unweighted projection but the 5th-95th uncertainty range of the weighted projection is 38% smaller with the reduction mainly in the upper bound, the largest reduction appears in the Southeast China.