Neural-network based approaches to understand regional climate change and climate predictability

Machine learning has become an increasingly common and useful tool within the broader fields of climate and earth science. In this talk, I will present two recent applications of machine learning to understand regional climate change and climate system predictability. First, I will discuss an example using machine learning to study changes in extreme precipitation. Global warming has increased the occurrence of extreme precipitation, but there is uncertainty around the physical mechanisms causing these changes, with debate about the role of changes in the atmospheric circulation. We use a convolutional neural network (CNN) to analyze circulation-driven versus thermodynamically-driven changes in U.S. Midwest extreme precipitation over the recent historical period. We find that the CNN can quickly learn the relationship between the large-scale circulation and extreme precipitation, with the CNN identifying most (>90%) of extreme precipitation days based on the large-scale circulation alone. Additionally, we find that atmospheric circulation conditions associated with extreme precipitation have become more frequent over the past 20 years, and that these conditions are more likely to result in extreme precipitation now than in the past because of increases in atmospheric moisture flux to the Midwest. Second, I will discuss recent research to uncover multi-year climate predictability using neural networks. For example, recent work within the field has demonstrated skillful neural network-based predictions of global surface temperature, including global warming slowdowns (Labe and Barnes, 2022), and decadal ocean variability, (e.g. Gordon and Barnes, 2022). However, because neural network training requires large amounts of data, many studies so far have focused on identifying sources of predictability within large ensemble climate model simulations. I will end the talk by highlighting a new research direction to understand how this predictability uncovered in climate models translates to the real climate system.