Amazon mega-flood naratives from large ensemble simulations – are they unseen or unrealistic?

Large ensemble simulations can be exploited to generate larger data samples than the observed record and consequently better assess the likelihood of rare events. Such simulations have the potential to inform storylines of ‘unseen’ flood episodes, i.e., that are more extreme than those seen in historical records. This method has, for example, been used to improve design levels of storm-surges in the river Rhine and to anticipate rainfall extremes over the UK. However, adequate evaluation of simulated ‘unseen’ events is a complex task.
Here, we showcase simulated Amazonian mega-floods from the combination of global climate model EC-Earth and global hydrological model PCR-GLOBWB. We introduce a three-step procedure to assess the realism of these mega-floods based on the model properties (step 1), as well as the statistical features (step 2) and physical credibility of the simulation (step 3). For the Amazon example, we find that the unseen floods were the result of an unrealistic bias correction of precipitation. 
We reflect on the different types of models that can be used to generate large sample sizes, and discuss the difference between storylines from large ensembles as compared to targeted model experiments to identify mega-floods. We conclude that understanding the driving mechanisms of unseen events may guide future research by uncovering key model deficiencies. They may also play a vital role in helping decision makers to anticipate unseen impacts by detecting plausible drivers.