Identifying the response of extreme precipitation to warming by using interpretable neural networks

Abstract:

Deep learning methods such as ANN have been widely considered as black boxes in the past. However, recent developments of interpretation techniques for ANN allow researchers to backtrack the predictions to examine the patterns ANN considers important for the predictability. To examine how ANN separates the forced climate signal in extreme precipitation from internal variability and inter-model variability, we deploy the ANN interpretation method layerwise relevance propagation. The ANN successfully identifies the time-varying signal from the noise in CMIP5 and CMIP6 models following a high-emissions scenario. At the end of the 21st century, drying in subtropics and wetting in African and Indian monsoon regions contribute to an increasing relevance for detecting the forced response. We use the different predictions by the ANN to identify inter-model variability and regions of importance for these predictions. We identify a subset of models with extended subtropical dry regions during the historical period, which resemble the expected climate change response, as opposed to a subset of models with a narrower dry region. Our results illustrate how interpretable ANNs can be useful to identify systematic inter-model variability and to identify possible observational constraints to reduce the model uncertainty.