Applying Explainable Neural Networks to Detect Oceanic Patterns contributing to PDO Prediction Skill
Abstract
Understanding oceanic variability is crucial to the success of decadal climate predictions. Among the most important processes driving this long-term ocean variability is the Pacific Decadal Oscillation (PDO), which is associated with both large-scale climate shifts and regional climate predictability. Here we demonstrate that neural networks can predict both phase transitions and persistence in the PDO on 1-2 year timescales in the CESM2 pre-industrial control. We further show, using explainable AI, that our neural network predictions can be linked to previously theorized mechanisms that contribute to PDO transitions, namely the build up of ocean heat content in the off-equatorial western Pacific 12-24 months before a transition. With the success of this method, we further show how oceanic variability contributes to the prediction of surface variables on decadal timescales across the globe, and use XAI to identify patterns more likely to lead to predictive skill (so-called forecasts of opportunity). These explainability techniques show how we can discover and investigate the mechanisms behind decadal predictability and advance our understanding of our chaotic climate system.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.A15E1671G