Disentangling forced and internal components in observed climate evolution using spatiotemporal deep learning

Abstract

Attribution of observed historical climate evolution to externally forced change and internal variability is fundamental to understanding climate change and evaluating climate models. Yet such attribution remains challenging because the forced response is intertwined with internal variability in observations. Here, we develop a convolutional long short-term memory network (ConvLSTM) to disentangle these two components. Trained and evaluated using large-ensemble simulations from multiple climate models, ConvLSTM consistently outperforms existing model-based, statistical, and machine-learning approaches. Applying ConvLSTM to observations reveals an evolving spatial pattern of the forced response and substantial contributions from internal variability to recent multidecadal climate trends. For several key climate features whose attribution remains debated, ConvLSTM indicates that internal variability amplified the forced reversal of Atlantic interhemispheric temperature asymmetry, dominated the observed La Niña-like warming pattern over the equatorial Pacific and the associated precipitation decline over southwestern North America since 1980, and modulated Arctic amplification around its nearly constant threefold forced value. Comparison between the ConvLSTM-estimated and model-simulated forced responses reveals broadly consistent patterns but also important discrepancies, with climate models simulating stronger Northern Hemisphere warming, weaker subtropical high intensification, and smaller land precipitation declines. Overall, ConvLSTM provides a robust framework for attributing historical climate evolution and evaluating the fidelity of climate models