Distribution-based pooling for combination of climate simulations
To study, assess and anticipate climate change, dozens of global climate models have been designed, each modeling the Earth system in a slightly different way. To extract a robust signal from the various simulations and outputs, the models are typically grouped into multi-model ensembles. These are then summarized in different ways, for example by multi-model means, medians or quantiles (possibly weighted). In this work, we introduce a new probability aggregation method called “alpha pooling” that constructs an aggregated distribution function (RF) designed to be closer to a reference RF over the calibration (historical) period. The aggregated RFs can then be used to adjust the bias of the raw climate simulations, thus performing a “multi-model bias correction”. The key to α-pooling is a parameter α that describes the type of transformation and thus the type of aggregation, generalizing both linear and log-linear pooling methods. We first establish that α-pooling is a mathematically valid aggregation method by verifying some optimal properties. Then, focusing on climate model simulations of temperature and precipitation over Western Europe, several experiments are conducted to evaluate the performance of the method.