AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-21
Disentangling forced and internal components in observed climate evolution using spatiotemporal deep learning
Attributing observed historical climate evolution to externally forced change versus internal variability is fundamental to understanding climate change and evaluating climate models, yet the two are intertwined in observations. The authors develop a convolutional LSTM (ConvLSTM), trained and evaluated on large-ensemble simulations from multiple climate models, that consistently outperforms existing model-based, statistical, and machine-learning approaches. Applied to observations, it reveals an evolving spatial pattern of the forced response and substantial internal-variability contributions to recent multidecadal trends—amplifying the forced reversal of Atlantic interhemispheric temperature asymmetry, dominating the La Niña-like equatorial Pacific warming and the precipitation decline over southwestern North America since 1980, and modulating Arctic amplification around its nearly constant threefold forced value—while comparison with model-simulated forced responses shows broadly consistent patterns alongside discrepancies such as stronger simulated Northern Hemisphere warming, weaker subtropical high intensification, and smaller land precipitation declines.