New classes of climate model emulators to improve paleoclimate reconstructions

Version 1 for publication "New classes of climate model emulators to improve paleoclimate reconstructions".

User guide :

This repository contains the code and processed datasets used to reproduce the results presented in the article “New classes of climate model emulators to improve paleoclimate reconstructions”.

The repository is organised into two main archives: `Code.zip` and `Data.zip`.

`Code.zip` contains the implementation of the dimensionality-reduction methods (EOF and autoencoder), the climate emulators (LIM, RC, RCn, AE, AERCn, and AERNN), and the notebooks used to reproduce the experiments and figures presented in the paper. The preprocessing of the IPSL-CM6A-LR simulations is performed in `Experiments/IPSL/Preprocessing_IPSL.ipynb`. This preprocessing is also the basis for the treatment of the CMIP6 simulations. The notebooks in `Experiments/IPSL/` reproduce the IPSL-based analyses, while the notebooks in `Experiments/CMIP6 models/` reproduce the multi-model analyses across CMIP6 simulations as well as the case studies for IPSL-CM6A-LR and MIROC6, including the ensemble prediction experiments. Several subdirectories in `Code.zip` contain their own README files describing the corresponding scripts, notebooks, and outputs in detail.

`Data.zip` contains the processed datasets and precomputed results used by the notebooks. In particular, it includes the processed IPSL datasets `tas_anom` (10 past2k ensemble members) and `tas_anom_novolc_2` (5 ensemble members without volcanic forcing), as well as the volcanic forcing time series `aerso_final`. The archive also contains emulator predictions for the IPSL-CM6A-LR simulation, evaluation metrics across CMIP6 models, pretrained models, and additional precomputed files used in the analyses. For the CMIP6 experiments, emulator training is repeated using increasing numbers of historical runs in order to analyse the sensitivity of emulator performance to training size. A dedicated README in `Data.zip` describes the structure and content of these files in detail.

To reproduce the main results of the paper, the recommended workflow is as follows. First, preprocess the IPSL simulations using `Experiments/IPSL/Preprocessing_IPSL.ipynb`. Then apply dimensionality reduction using the scripts in `Dimension_reduction/`. Next, train the emulators using the code in `Emulators/`. The emulator performance for IPSL can then be evaluated with `Experiments/IPSL/evaluate_emulators.ipynb`. The CMIP6 analyses can be reproduced using the notebooks in `Experiments/CMIP6 models/`.

The original climate model simulations used in this work are publicly available through the CMIP6 archive via the ESGF data portals.