Quantifying the Internal Variability of Midlatitude Storms Using Deep Learning

Extratropical storms shape midlatitude weather and are influenced by both the slowly evolving climate and rapid changes in synoptic conditions. While the impact of each factor has been extensively studied, their relative importance remains uncertain, creating challenges in resolving the signal-to-noise ratio necessary for attributing extratropical weather events to current anthropogenic climate change. Here, we quantify the climate's relative importance in both climatic storm activity and individual storm development using 84 years of ERA-5 data, tracks of 100,00 cyclones and 50,000 anticyclones, and Convolutional Neural Networks (CNNs). We find that the constructed CNN model predicts over 90% of the variability in climatic storm activity, indicating that, from a climatic perspective, internal variability is negligible. In contrast, a similar model predicts less than one-third of the variability in individual storm features, such as intensity, growth time, and trajectory, demonstrating that their variability is dominated by internal variability. Using this estimate of internal variability and the mean impact of present-day climate change, we calculate a signal-to-noise ratio for attribution of storm intensity of approximately 0.2%, highlighting the significant challenge in attributing extreme individual storms to anthropogenic climate change. However, a signal-to-noise ratio ten times higher is obtained for warm heat anomalies associated with storms, emphasizing the potential for attributing storm-related impacts that are directly linked to climate change.