OHW22 ENSO Prediction

Summary

We propose a project to develop a framework for ENSO prediction using a range of machine learning approaches and benchmark their skill against other methods. Identify observables that are most relevant for ENSO predictability at a particular timescale.

Data description

• Coarsened CESM2 Large Ensemble members (SST fields) from 1° to 3° resolution
• Used the first 10 of 100 CESM2 LE members

Deep learning components

• 2D CNN (relu/selu, adam/nadam/adamx)
• 3D CNN (relu/selu, adam/nadam/adamx)
• EOF/LSTM
• CNN/LSTM
• Transformers?
• Create a baseline method to compare these methods to (i.e. persistence method, AR1 model)
• Interpretability– why do the results look this way?

Opportunities for growth

• EOF-LSTM: Instead of EOFs, we can use a CNN to predict one number (or however number modes we want to keep). This is a more complicated dimension reduction than PCA and then use LSTM. So CNN then LSTM as the workflow.