Predicting climate scale oceanic processes: leveraging uncertainty quantification and XAI to increase trust and identify predictability
Abstract
We discuss techniques for generating trustworthy neural network predictions of climate scale ocean variability. Firstly, we demonstrate how incorporating a confidence output into a neural network's prediction can increase trust in the network. By isolating the most confident predictions as identified by the neural networks, we examine which samples the neural network views as predictable. For example, we examine predictable SST evolution in the North Atlantic Subpolar Gyre on decadal timescales, as well as predictable persistence and transitions in the Pacific Decadal Oscillation. In these applications, we show that the confident predictions appear to be grounded in current theories for predictability of decadal scale climate processes. Secondly, we use explainable AI to examine the inputs that are relevant to these confident predictions and match the findings by the neural networks to the current understanding of mechanisms behind decadal prediction and predictability. We demonstrate that relevant regions for predictability identified by the networks both agree with our current understanding of large scale ocean variability, and suggest new avenues to explore.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMOS35A..05G