We use neural networks and large climate model ensembles to explore predictability of internal variability in sea surface temperature anomalies on interannual (1-3 year) and decadal (1-5 and 3-7 year) timescales. We find that neural networks can skillfully predict SST anomalies at these lead times, especially in the North Atlantic, North Pacific, Tropical Pacific, Tropical Atlantic and Southern Ocean. The spatial patterns of SST predictability vary across the nine climate models studied. The neural networks identify “windows of opportunity” where future SST anomalies can be predicted with more certainty. Neural networks trained on climate models also make skillful SST predictions in reconstructed observations, although the skill varies depending on which climate model the network was trained. Our results highlight that neural networks can identify predictable internal variability within existing climate datasets and show important differences in how well patterns of SST predictability in climate models translate to the real world.

Improving predictions of decadal climate variability is critical for reducing uncertainty in near-term climate change. Here we investigate the potential to improve prediction skill in the North Pacific by identifying predictable patterns of sea surface temperatures (SSTs) in climate simulations, and then applying them to observations. A convolutional neural network (CNN) is first trained to predict basin-wide SSTs in the North Pacific on 1-5 year time-scales in nine global climate models (GCMs), and a pattern of high skill is identified from the GCM data. This pattern of high skill learned from GCMs is then skillfully predicted by the CNN when given observations as inputs. The identified pattern is notably not the Pacific Decadal Oscillation, and instead corresponds to basinwide warming and cooling focused in the North Pacific Gyre. We conclude that investigating the mechanisms that contribute to predictability (rather than variability) is an effective avenue for improving near-term climate predictions.

Predictable internal climate variability on decadal timescales (2-10 years) is associated with large-scale oceanic processes, however these predictable signals may be masked by the noisy climate system. One approach to overcoming this problem is investigating state-dependent predictability - how differences in prediction skill depend on the initial state of the system. We present a machine learning approach to identify state-dependent predictability on decadal timescales in the Community Earth System Model version 2 by incorporating uncertainty estimates into a regression neural network. We leverage the network’s prediction of uncertainty to examine state dependent predictability in sea surface temperatures by focusing on predictions with the lowest uncertainty outputs. In particular, we study two regions of the global ocean - the North Atlantic and North Pacific - and find that skillful initial states identified by the neural network correspond to particular phases of Atlantic multi-decadal variability and the interdecadal Pacific oscillation.

Predicting Pacific Decadal Oscillation (PDO) transitions and understanding the associated mechanisms has proven a critical but challenging task in climate science. As a form of decadal variability, the PDO is associated with both large-scale climate shifts and regional climate predictability. We show that artificial neural networks (ANNs) predict PDO persistence and transitions from 12 months onward. Using layer-wise relevance propagation to investigate the ANN predictions, we demonstrate that the ANNs utilize oceanic patterns that have been previously linked to predictable PDO behavior. For PDO transitions, ANNs recognize a build-up of ocean heat content in the off-equatorial western Pacific 12-27 months before a transition occurs. The results support the continued use of ANNs in climate studies where explainability tools can assist in mechanistic understanding of the climate system.