Sensitivity of extreme precipitation to climate change inferred using artificial intelligence shows high spatial variability

Abstract

Abstract Evaluating how extreme precipitation changes with climate is challenged by the paucity, brevity and inhomogeneity of observational records. Even when aggregating precipitation observations over large regions (obscuring potentially important spatial heterogeneity) the statistics describing extreme precipitation are often too uncertain to be of any practical value. Here we present an approach where a convolutional neural network (an artificial intelligence model) is trained on precipitation measurements from 10,000 stations to learn the spatial structure of the parameters of a generalised extreme value model, and the sensitivity of those parameters to the annual mean, global mean, surface temperature. The method is robust against the limitations of the observational record and avoids the short-comings of regional and global climate models in simulating the sensitivity of extreme precipitation to climate change. The maps of the sensitivity of extreme precipitation to climate change, on ~1.5 km × 1.5 km grids over North America, Europe, Australia and New Zealand, derived using the successfully trained convolutional neural network, show high spatial variability.