Seasonal-level prediction of climate data with a variational autoencoder

Overview

Jupyter notebooks for the implementation of a variational autoencoder (VAE) for climate data modeling and prediction.

The Jupyter notebooks demonstrate the process of training and exploring a Variational Autoencoder (VAE) on precipitation and sea-surface temperature data. The training process is divided into two steps: pre-training on CMIP6 data and transfer learning on observational data.

Requirements

1. The Jupyter notebooks requires the VAE package, which is available at:

   https://github.com/andr-groth/VAE-project

2. Sample data used in the notebook is included in the data/ folder. The data is in netCDF format and has been prepared with the help of the CDO scripts, which are available at:

   https://andr-groth.github.io/CDO-scripts

   For more information on the data preparation see data/README.md.

Usage

This repository contains two Jupyter notebooks for model training and prediction:

1. To train the VAE on CMIP6 data and transfer learn on observational data: VAEp_train.ipynb

2. To make predictions on observational data: VAEp_explore.ipynb

The notebooks use sample data provided in the data/ folder in this repository.