Learning algorithms allow for improved reliability and accuracy of global mean surface temperature projections

Climate predictions are only meaningful if the associated uncertainty is reliably estimated. A standard practice is to use an ensemble of climate model projections. The main drawbacks of this approach are the fact that there is no guarantee that the ensemble projections adequately sample the possible future climate conditions. Here, we suggest using simulations and measurements of past conditions in order to study both the performance of the ensemble members and the relation between the ensemble spread and the uncertainties associated with their predictions. Using an ensemble of CMIP5 long-term climate projections that was weighted according to a sequential learning algorithm and whose spread was linked to the range of past measurements, we find considerably reduced uncertainty ranges for the projected global mean surface temperature. The results suggest that by employing advanced ensemble methods and using past information, it is possible to provide more reliable and accurate climate projections. The ensemble spread of climate models is often interpreted as the uncertainty of the projection, but this is not always justified. Applying learning algorithms to an ensemble of climate predictions allows for a significant uncertainty reduction of projected global mean surface temperatures compared to the ensemble spread.