A Storyline Approach to Explore Impacts of Regional No-Return Temperature Thresholds Using Transfer Learning and Uncertainty Quantification of Explainable Neural Network Predictions

Disparities in both the magnitude and rate of warming throughout the globe have led to uncertainties in the timing of crossing critical heat thresholds. These thresholds are often relative to global annual-mean temperature anomalies, which can result in a loss of regional and seasonal information for climate change impacts. Here, convolutional neural networks are trained to predict the year in which a region will exceed, and not fall below (e.g. no-return) specific temperature thresholds for summertime minimum temperature, with uncertainties quantified for each prediction using multiple CMIP6 climate projection scenarios. Transfer learning is used to update networks with additional training from observations to constrain the climate model projections to those which most align with the observations-based predictions. Explainability techniques (i.e. XAI) applied here are integral to determining the large-scale global climate patterns and regions which have the highest contributions to determining the timeline of thresholds being crossed, offering insight for regional impacts and predictability changes from global climate change. A storyline approach is implemented, in which individual ensemble members closely resembling the network predictions are selected for investigation of the specific regional impacts due to crossing the temperature thresholds. XAI heatmaps for single ensemble members pinpoint climate patterns that impact various regional summertime minimum temperatures under different climate change scenarios. Performing this analysis within single ensemble members allows for assessment of regional impacts of climate change in physically-viable scenarios without losing information from ensemble means.