To what extent can we predict the climate impacts of internal climate variability on decadal timescales? Recent research has suggested the utility of large ensembles of fully coupled climate simulations for this grand challenge. Efforts to improve our documentation and physical understanding of decadal variability and predictability have accelerated as large ensembles of fully coupled climate simulations have become more readily available. Moreover, ongoing advancements in model fidelity and ensemble size have improved the capabilities of climate models to produce accurate representations of the real-world climate system, leading to an expectation of improved skill in decadal predictions.

We use neural networks to identify regimes of sea-surface temperature (SST) anomalies that lead to predictable patterns in near-surface temperature and precipitation on decadal timescales within the CESM Large Ensemble (CESM-LE). The neural networks are trained to identify and localize SST regimes that evolve in tandem with decadal anomalies in temperature and precipitation in specific regions of the world, such as the southwestern United States. In doing so, we isolate the particular SST regimes within CESM-LE that can be used for prediction on decadal timescales. These results have implications for improving the skill of decadal prediction systems, particularly if the favorable SST regimes within CESM-LE can be similarly identified within observational data sets or decadal prediction ensembles, such as the CESM Decadal Predictability Large Ensemble (CESM-DPLE).

Abstract Author(s)
Benjamin A. Toms, Elizabeth A. Barnes, Jim W. Hurrell
University
Colorado State University