Climsim_test

This repo uses deep learning to improve climate parameterization. The dataset used in this repo is ClimSim: https://arxiv.org/abs/2306.08754

0. Preprocessing: climsim_data.ipynb

This notebook provides the details about how to get the data from the raw .nc files. This part I only used the last-year data (val_input.npy and val_target.npy).

1. Data Downloading

All of the analyses are based on the low-resolution, real-geography dataset. To download the input and output variables for the training and validation sets, go to: https://huggingface.co/datasets/LEAP/subsampled_low_res/tree/main. Download train_input.npy, train_target.npy, val_input.npy, val_target.npy. Or execute download_data.ipynb to download data from Huggingface directly. The normalization and scaling files can be found at: https://github.com/leap-stc/ClimSim/tree/main/preprocessing/normalizations These .nc files are required for post processing.

2. Baseline model: FCNN.ipynb

Use a one-layer NN to train and test. This will generate metric files for this model. Files will be stored in the metrics and metrics_netcdf folders.

3. Baseline model: CNN.ipynb

Unfinished.

3. Baseline model: MLP.ipynb

Finished. Metrics are slightly different from paper due to the unknown training epochs.

4. Baseline model: transformer_test.ipynb

Test transformer.

5. reshape time space.ipynb

Test the reshaping of the dataset for transformer. Can be deleted.

Use test_model.ipynb to test a new model (a simple NN is already given in this file). Do not use quickstart_example.ipynb since it does not provide the actual values for MAE/RMSE/R2.