An interpretable neural network approach to identifying sources of predictability in the future climate

Earth system predictability on decadal timescales can arise from both low frequency internal variability as well as from anthropogenically forced long-term changes. However, on these timescales, the chaotic nature of the climate system makes skillful predictions difficult to achieve even if we include information from climate change projections. Furthermore, it is difficult to separate the contributions from internal variability and external forcing to predictability. One way to improve skill is through identifying and harnessing initial conditions with more predictable evolution, so-called state-dependent predictability. We explore a neural network approach to identify these opportunistic initial states in the CESM2 large ensemble and subsequently explore how predictability may manifest in a future climate, influenced by both forced warming and internal variability. We use an interpretable neural network to demonstrate that internal variability will continue to play an important role in future climate predictions, especially for states of increased predictability.