Distinguishing the Regional Emergence of United States Summer Temperatures Between Observations and Climate Model Large Ensembles

Despite the rapid rise in global temperatures, it can be difficult at times to distinguish externally-forced climate signals from internal variability at smaller spatial and temporal scales. Given the large variability in the rate of warming across the contiguous United States (CONUS) in summer, we leverage a newly designed artificial neural network (ANN) architecture to further examine the regional timing of emergence (ToE) of mean maximum and minimum temperatures. The ToE metric is particularly useful for societal and ecological planning, as it quantifies when changes in a given climate variable have exceeded the historical range of natural variability. We first train our ANN on CONUS maps of summer temperatures using simulations from climate model large ensembles and then task the ANN with outputting the year of each map. This can then be translated into a year when the pattern of forced climate change is first detectable (i.e., the ToE). To identify the important regions of temperature change, we complement the ToE predictions with the use of explainable artificial intelligence (XAI) techniques, such as layer-wise relevance propagation and integrated gradients.

The earliest ToE is detected for seasonally-averaged maps of minimum temperature, but in contrast, we do not find any emergence for summer maximum temperatures outside of the western CONUS. Interestingly, we find an improvement in ANN skill after training on higher resolution climate model data, especially for predicting the year of temperature maps in the 20th century. We compare composites of XAI maps to understand how the ANN can still make accurate predictions when there is substantially weaker radiative forcing and subsequently find that the high 20th century skill is associated with temperatures in the vicinity of topography and prescribed land-use/land-cover change. This work suggests that the choice of horizontal resolution in training data may be an important consideration for applications of machine learning in climate science.