PNNL

Abstract

Observed Arctic amplification (AA), defined as the ratio of Arctic to global-mean surface air temperature (SAT) trends, peaks in April—a feature not captured by climate models. Here, we quantify the impact of internal variability on Arctic SAT trends in March, April and May from 1980 to 2022 using multiple approaches, including machine learning techniques—convolutional neural networks (CNNs) and partial least squares (ML-PLS)—as well as dynamical adjustment and wind-nudged experiments. Across all approaches, internal variability emerges as a substantial contributor to Arctic warming in these months, but its influence is strongly monthly dependent. April is particularly prominent: internal variability accounts for ~40% of the observed Arctic SAT trend, a result consistent between the CNN and wind-nudged estimates. For each spring month, the linear methods (ML-PLS and dynamical adjustment) yield systematically smaller contributions than the CNN, indicating that approaches capable of capturing nonlinear relationships more effectively isolate the internal variability component. Removing internal variability largely reconciles the springtime monthly AA discrepancy between models and observations. The strong spatial correlations between the difference of SAT trend pattern (observations minus the multi-model mean) and that derived from wind-nudged experiments further implies that the simulated forced trend pattern is broadly realistic. Finally, the observed springtime monthly circulation–temperature configurations are rare in coupled climate models across all three spring months, suggesting that models have difficulty reproducing the magnitude of Arctic internal variability.