A Data-Driven Approach to Identifying Key Regions of Change Associated with Future Climate Scenarios

One of the largest sources of uncertainty in climate change projections is related to the evolution of future greenhouse gas emissions. This is otherwise known as scenario uncertainty and is typically accounted for by considering different versions of the future, such as through the Shared Socioeconomic Pathways (SSPs) that are applied to the latest generation of global climate models. However, there has generally been less work on understanding the climate response to alternative 21st century mitigation scenarios, such as overshoot pathways which leverage technology like carbon capture and storage to induce negative emissions. In addition, it also remains unclear how to best track which SSP scenario is most closely aligned with real-world observations. To address these open questions, we leverage recent advances in explainable artificial intelligence (XAI) to introduce a detection method for linking maps of climate variables to individual SSP scenarios. Specifically, we train an artificial neural network (ANN) on annual mean global climate maps using output from a collection of different historical, natural, and future forcing scenarios from the Geophysical Fluid Dynamics Laboratory’s Seamless System for Prediction and EArth System Research (SPEAR). We task the ANN to classify which forcing scenario is associated with each temperature or precipitation map and then utilize ad hoc feature attribution XAI tools to understand how the ANN is learning to make its correct classifications. After training the ANN on output from these SPEAR simulations, we then test out-of-sample climate data from two overshoot scenarios conducted with SPEAR where aggressive climate mitigation begins in either 2040 (SSP5-3.4OS) or 10 years earlier. The XAI methods for these overshoot scenarios thus reveal important regions of change, such as over the North Atlantic and Central Africa, that are associated with the response to mitigation occurring at relatively slower or faster timescales. In addition to quantifying the benefits of rapidly reducing carbon emissions, this data-driven approach can learn to associate the impacts of global climate change with different historical or future emission scenarios for monitoring in near-real time.