Transferring climate change physical knowledge

Repository including the code needed to reproduce the results present in Immorlano et al., Transferring climate change physical knowledge, PNAS, Volume 122, Issue 14, 2025.

Content

• Related Publication
• Installation
• Run a demo version
• Run the full version
• Reproduce the results present in the paper
• Contributors
• Acknowledgements and References
• License

Related Publication

F. Immorlano, V. Eyring, T. le Monnier de Gouville, G. Accarino, D. Elia, S. Mandt, G. Aloisio, P. Gentine, Transferring climate change physical knowledge, Proc. Natl. Acad. Sci. U.S.A. 122 (15) e2413503122, https://doi.org/10.1073/pnas.2413503122 (2025).

Installation

Python version 3.11.0 or higher is needed.

A conda env containing all the packages and versions required to run the workflow and/or reproduce the figures present in the paper can be created by running the following command:

conda env create --file transferring_env.yml

This makes the installation easy and fast. The conda env was created and tested on a MacBook M2 Pro with MacOS Ventura 13.4.1 to enable the user to run a demo version of the workflow and reproduce the results present in the paper.

Concerning the full version, Tensorflow version 2.12.0 was used to run the entire workflow on Ginsburg (Columbia University) and Juno (CMCC Foundation) supercomputers.

Files description

• architectures.py → Definition and building of the Deep Neural Network (DNN) used in the Transfer Learning (TL) approach
• area_cella.csv → Area weighting factors for each gridpoint of the grid corresponding to CanESM5_CanOE_grid used in this work
• CanESM5_CanOE_grid → Description of the grid used in this work (Grid from CanESM5-CanOE simulation)
• transferring_env.yml → YAML file needed to create the conda environment to reproduce the experiments and the figures
• erf_estimates_with_aerosols_Zebedee_Nichols.csv → Dataset of Effective Radiative Forcing values for each year
• lib.py → Routines called in several scripts
• variables.py → Definition of variables used in the scripts
• BEST_regridded_annual_1979-2022.nc → Berkeley Earth Surface Temperatures (BEST) observational maps resulting from the preprocessing (i.e., regridding to CanESM5-CanOE grid, computation of annual average maps, deletion of years out of 1979-2022)
• gaussian_noise_5 → Directory containing BEST observational maps added with Gaussian noise to be used for TL on observations
• Land_and_Ocean_global_average_annual.txt → Average annual uncertainties associated with BEST observational data. The column headers were manually removed for ease of use

The following scripts are used to download and process CMIP6 and BEST data:

• CMIP6_download_process.py → Download of CMIP6 simulations from Climate Data Store and processing
• BEST_data_processing.py → Processing of BEST observational maps
• BEST_data_add_gaussian_noise.py → Generation of BEST observational maps added with Gaussian noise to be used for TL on observations

The following scripts are used to perform Training, TL on Simulations (leave-one-out cross-validation procedure) and the TL on Observations:

• First_Training.py → Training of each DNN on one of the 22 CMIP6 Earth System Models (ESMs) simulations under one of the three SSP scenarios (SSP2-4.5, SSP3-7.0, SSP5-8.5)
• Transfer_learning_simulations.py → TL of the (pre-trained) DNNs on ESMs simulations according to the leave-one-out cross-validation (LOO-CV) procedure
• Transfer_learning_observations.py → TL of the (pre-trained) DNNs on BEST observational data

The following scripts must be used to reproduce the figures and compute the values for Supplementary Table 1 present in the paper:

• Fig_1.py
• Fig_2.py
• Fig_3.py
• Fig_4.py
• Fig_S1.py
• Fig_S2.py
• Fig_S3.py
• Fig_S4.py
• Fig_S5.py
• Fig_S6.py
• Fig_S7.py
• Fig_S8.py
• Fig_S9.py
• Fig_S10.py
• Table_S2.py
• Table_S3.py
• Table_S4.py
• Table_S5.py
• Table_S6.py
• Uncertainty_decomposition.py

Run a demo version

A demo version fo the entire workflow can be run. Once the GitHub repository has been downloaded, the files required to run the demo must be organised according to the following hierarchy:

├── root
│   ├── architectures.py
│   ├── area_cella.csv
│   ├── BEST_data_add_gaussian_noise.py
│   ├── BEST_data_processing.py
│   ├── CanESM5-CanOE_grid
│   ├── CMIP6_download_process.py
│   ├── Demo_download
│   │   ├── Data
│   │   │   ├── BEST_data
│   │   │   │   ├── Land_and_Ocean_global_average_annual.txt
│   ├── Demo_no_download
│   │   ├── Data
│   │   │   ├── BEST_data
│   │   │   │   ├── BEST_regridded_annual_1979-2022.nc
│   │   │   │   ├── gaussian_noise_5
│   │   │   │   ├── Land_and_Ocean_global_average_annual.txt
│   │   │   ├── CMIP6_data
│   ├── erf_estimates_with_aerosols_Zebedee_Nichols.csv
│   ├── Figures_and_Tables
│   ├── First_Training.py
│   ├── lib.py
│   ├── Transfer_learning_observations.py
│   ├── Transfer_learning_simulations.py
│   ├── variables.py

The demo version can be run in two ways:

• download and create a small dataset and then train the DNNs on those data
• start directly with training the DNNs on the same small dataset already downloaded and processed (present in Demo_no_download)

In both cases, the workflow will be executed for CNRM-ESM2-1, FGOALS-f3-L and MIROC6 ESMs simulations under SSP2-4.5.

Data downloading and processing + training

To this aim, the BEST observational maps used in this study (i.e., Global Monthly Land + Ocean — Average Temperature with Air Temperatures at Sea Ice (Recommended; 1850 – Recent) — 1º x 1º Latitude-Longitude Grid) must be gathered from the BEST archive (direct download: 1º x 1º Latitude-Longitude Grid (~400 MB)) . The file must be named BEST_2022.nc and placed in root/Demo_download/Data/BEST_data.
Since CMIP6 data were gathered from the Copernicus Climate Data Store (CDS), the CDS API key should be configured on your laptop according to the CDS API documentation before downloading them.
The following variables in variables.py should be set to:

• demo_download = True
• demo_no_download = False

Now, the scripts should be executed in the following order:

1. CMIP6_download_process.py to download and process (i.e., regridding to CanESM5-CanOE grid, computing annual average maps) CMIP6 simulation data for the three ESMs and for SSP2-4.5. The resulting nc files will be saved in root/Demo_download/Data/CMIP6_data/near_surface_air_temperature/Annual_uniform_remapped.
2. BEST_data_processing.py to process BEST observational maps (i.e., regridding to CanESM5-CanOE grid, computing annual average maps, deleting years out of 1979-2022). The resulting file will be saved as root/Demo_download/Data/BEST_data/BEST_regridded_annual_1979-2022.nc
3. BEST_data_add_gaussian_noise.py to add noise to the BEST observational maps. Specifically, the noise is sampled from a Gaussian distribution with 0 mean and stddev equal to the BEST observational data uncertainty. The addition of noise is repeated 5 times for each ESM and each scenario. The resulting files will be saved in root/Demo_download/Data/BEST_data/gaussian_noise_5.
4. First_Training.py to train an individual DNN on the simulation of one of the three ESMs for the SSP2-4.5 scenario. A total of three DNNs are trained, each with the same architecture. The results will be saved in root/Demo_download/Experiments/First_Training/First_Training_YYYY-MM-DD_hh-mm-ss.
5. Transfer_learning_simulations.py to load the pre-trained DNNs and fine tune them on the ESMs simulations according to the LOO-CV procedure. The variable FIRST_TRAINING_DIRECTORY in Transfer_learning_simulations.py must be set to the directory name related to the first training (i.e. FIRST_TRAINING_DIRECTORY = First_Training_YYYY-MM-DD_hh-mm-ss). This is necessary to load the pre-trained DNNs. The variable exclude_family_members in variables.pymust be set to True if the ESMs based on the same atmospheric model as the take-out ESM must be excluded in each iteration of the LOO-CV procedure. The results will be saved in root/Demo_download/Experiments/Transfer_Learning_on_Simulations/Transfer_learning_YYYY-MM-DD_hh-mm-ss/Shuffle_[number]. The Shuffle_[number] corresponds to an iteration of the LOO-CV procedure.
6. Transfer_learning_observations.py to load the pre-trained DNNs and fine tune them on the BEST observational data. The variable FIRST_TRAINING_DIRECTORY in Transfer_learning_observations.py must be set to the directory name related to the first training (i.e. FIRST_TRAINING_DIRECTORY = First_Training_YYYY-MM-DD_hh-mm-ss). This is necessary to load the pre-trained DNNs. The results will be saved in root/Demo_download/Experiments/Transfer_Learning_on_Observations/Transfer_learning_obs_YYYY-MM-DD_hh-mm-ss.

Training (without downloading and processing)

A demo version can be run directly starting with training the DNNs on the CNRM-ESM2-1, FGOALS-f3-L and MIROC6 models simulations.
In this case, the following variables in variables.py should be set to:

• demo_download = False
• demo_no_download = True

Now, the scripts should be executed in the following order:

1. First_Training.py to train an individual DNN on one of the three ESMs simulation for the SSP2-4.5 scenario. A total of three DNNs will be trained, each with the same architecture. The results will be saved in root/Demo_no_download/Experiments/First_Training/First_Training_YYYY-MM-DD_hh-mm-ss.
2. Transfer_learning_simulations.py to load pre-trained DNNs and fine tune them on the ESMs simulations according to the LOO-CV procedure. The variable FIRST_TRAINING_DIRECTORY in Transfer_learning_simulations.py must be set to the directory name related to the first training (i.e. FIRST_TRAINING_DIRECTORY = First_Training_YYYY-MM-DD_hh-mm-ss). This is necessary to load the pre-trained DNNs. The variable exclude_family_members in variables.py must be set to True if the ESMs based on the same atmospheric model as the take-out ESM must be excluded in each iteration of the LOO-CV procedure. The results will be saved in root/Demo_no_download/Experiments/Transfer_Learning_on_Simulations/Transfer_learning_YYYY-MM-DD_hh-mm-ss/Shuffle_[number]. The Shuffle_[number] corresponds to an iteration of the LOO-CV procedure.
3. Transfer_learning_observations.py to load the pre-trained DNNs and fine tune them on the BEST observational data. The variable FIRST_TRAINING_DIRECTORY in Transfer_learning_observations.py must be set to the directory name related to the first training (i.e. FIRST_TRAINING_DIRECTORY = First_Training_YYYY-MM-DD_hh-mm-ss). This is necessary to load the pre-trained DNNs. The results will be saved in root/Demo_no_download/Experiments/Transfer_Learning_on_Observations/Transfer_learning_obs_YYYY-MM-DD_hh-mm-ss.

The Demo software was tested on a MacBook M2 Pro equipped with MacOS Ventura 13.4.1. The expected run times are the following:

• CMIP6_download_process.py: about 12 seconds (after the CMIP6 data download)
• BEST_data_processing.py: about 122 seconds
• BEST_data_add_gaussian_noise.py: about 7 seconds
• First_Training.py, Transfer_learning_simulations.py, Transfer_learning_observations.py: about 4.5 seconds per epoch

Run the full version

The full version of the entire workflow can be run. Once BEST_data.zip and CMIP6_data.zip have been downloaded from Zenodo, the files needed to run the entire workflow must be organised according to the following hierarchy:

├── root
│   ├── architectures.py
│   ├── area_cella.csv
│   ├── BEST_data_add_gaussian_noise.py
│   ├── BEST_data_processing.py
│   ├── CanESM5-CanOE_grid
│   ├── CMIP6_download_process.py
│   ├── erf_estimates_with_aerosols_Zebedee_Nichols.csv
│   ├── First_Training.py
│   ├── lib.py
│   ├── Source_data
│   │   ├── BEST_data
│   │   │   ├── BEST_regridded_annual_1979-2022.nc
│   │   │   ├── gaussian_noise_5
│   │   │   ├── Land_and_Ocean_global_average_annual.txt
│   │   ├── CMIP6_data
│   ├── variables.py

The full version will be run by training the DNNs on the 22 ESMs simulations for SSPs 2-4.5, 3-7.0, and 5-8.5.
In this case, the following variables in variables.py must be set to:

• demo_download = False
• demo_no_download = False

Now, the scripts should be executed in the following order:

1. First_Training.py to train an individual DNN on one of the 22 ESMs simulation for the SSPs 2-4.5, 3-7.0, and 5-8.5. A total of 66 DNNs will be trained, each with the same architecture. The results will be saved in root/Experiments/First_Training/First_Training_YYY-MM-DD_hh-mm-ss.
2. Transfer_learning_simulations.py to load the pre-trained DNNs and fine tune them on the ESMs simulations according to the LOO-CV procedure. The variable FIRST_TRAINING_DIRECTORY in Transfer_learning_simulations.py must be set to the directory name related to the first training (i.e. FIRST_TRAINING_DIRECTORY = First_Training_YYY-MM-DD_hh-mm-ss). This is necessary to load the pre-trained models. The variable exclude_family_members in variables.py must be set True if the ESMs based on the same atmospheric model as the take-out ESM must be excluded in each iteration of the LOO-CV procedure. The results will be saved in root/Experiments/Transfer_Learning_on_Simulations/Transfer_learning_YYY-MM-DD_hh-mm-ss/Shuffle_[number]. The Shuffle_[number] corresponds to an iteration of the LOO-CV approach.
3. Transfer_learning_observations.py to load the pre-trained DNNs and fine tune them on the BEST observational data. The variable FIRST_TRAINING_DIRECTORY in Transfer_learning_observations.py must be set to the directory name of the first training (i.e. FIRST_TRAINING_DIRECTORY = First_Training_YYY-MM-DD_hh-mm-ss). This is necessary to load the pre-trained models. The results will be saved in root/Experiments/Transfer_Learning_on_Observations/Transfer_learning_obs_YYY-MM-DD_hh-mm-ss.

Reproduce the results present in the paper

The figures and the results present in the paper can be reproduced. The variable compute_figures_tables in variables.py must be set to True. After having downloaded BEST_data.zip, CMIP6_data.zip, First_training_obs.zip, First_Training.zip, Transfer_Learning_on_Observations.zip, Transfer_Learning_on_Simulations_AM_families.zip, Transfer_Learning_on_Simulations_reverse.zip, and Transfer_Learning_on_Simulations.zip from Zenodo, the files needed to reproduce the results shall be organized as the following hierarchy:

├── root
│   ├── area_cella.csv
│   ├── Figures
│   │   ├── Fig_S1.py
│   │   ├── Fig_S2.py
│   │   ├── Fig_S3.py
│   │   ├── Fig_S4.py
│   │   ├── Fig_S5.py
│   │   ├── Fig_S6.py
│   │   ├── Fig_S7.py
│   │   ├── Fig_S8.py
│   │   ├── Fig_S9.py
│   │   ├── Fig_S10.py
│   │   ├── Fig_1.py
│   │   ├── Fig_2.py
│   │   ├── Fig_3.py
│   │   ├── Fig_4.py
│   │   ├── lats.pickle
│   │   ├── lons.pickle
│   │   ├── Table_S2.py
│   │   ├── Table_S3.py
│   │   ├── Table_S4.py
│   │   ├── Table_S5.py
│   │   ├── Table_S6.py
│   │   ├── Uncertainty_partitioning.py
│   ├── Source_data
│   │   ├── BEST_data
│   │   ├── CMIP6_data
│   │   ├── First_Training_obs
│   │   ├── First_Training
│   │   ├── Transfer_Learning_on_Observations
│   │   ├── Transfer_Learning_on_Simulations_AM_families
│   │   ├── Transfer_Learning_on_Simulations_reverse
│   │   ├── Transfer_Learning_on_Simulations

• First_Training_obs → Near-surface air temperature maps generated by the DNNs solely trained on the BEST observational data from 1979 to 2022
• First_Training → Near-surface air temperature maps generated by the DNNs trained on the 22 ESMs simulations under SSP2-4.5, 3-7.0, and 5-8.5 scenarios
• Transfer_Learning_on_Observations → Near-surface air temperature maps generated by the DNNs after transfer learning on BEST observational data
• Transfer_Learning_on_Simulations_AM_families → Near-surface air temperature maps generated by the DNNs after transfer learning on the 22 ESMs simulations under SSPs 2-4.5, 3-7.0, and 5-8.5 (LOO-CV procedure). Those ESMs based on the same atmospheric model as the taken-out ESM (i.e., belonging to the same atmospheric model family) were excluded in each iteration of the LOO-CV procedure
• Transfer_Learning_on_Simulations_reverse → Near-surface air temperature maps generated by the DNNs after the reverse LOO-CV procedure. The pre-trained DNNs were transfer learned on the 22 ESMs simulations under SSP2-4.5, 3-7.0, and 5-8.5 scenarios, by using 2023–2098 as training set and 1850–2022 as test set
• Transfer_Learning_on_Simulations → Near-surface air temperature maps generated by the DNNs after transfer learning on the 22 ESMs simulations under SSPs 2-4.5, 3-7.0, and 5-8.5 (LOO-CV procedure)

The following scripts should be used to reproduce the figures and the tables present in the paper:

• Fig_1.py
• Fig_2.py
• Fig_3.py
• Fig_4.py
• Fig_S2.py
• Fig_S3.py
• Fig_S4.py
• Fig_S5.py
• Fig_S6.py
• Fig_S7.py
• Fig_S8.py
• Fig_S9.py
• Fig_S10.py
• Table_S2.py
• Table_S3.py
• Table_S4.py
• Table_S5.py
• Table_S6.py
• Uncertainty_partitioning.py

Contributors

• Francesco Immorlano (francesco.immorlano@cmcc.it)
• Veronika Eyring (veronika.eyring@dlr.de)
• Thomas le Monnier de Gouville (thomas.le-monnier-de-gouville@polytechnique.edu)
• Gabriele Accarino (gabriele.accarino@cmcc.it)
• Donatello Elia (donatello.elia@cmcc.it)
• Stephan Mandt (mand@uci.edu)
• Giovanni Aloisio (giovanni.aloisio@cmcc.it)
• Pierre Gentine (pg2328@columbia.edu)

Acknowledgements and References

If you use the resource in your research, please cite our paper:

@article{
      doi:10.1073/pnas.2413503122,
      author = {Francesco Immorlano and Veronika Eyring and Thomas le Monnier de Gouville and Gabriele Accarino and Donatello Elia and Stephan Mandt and Giovanni Aloisio and Pierre Gentine},
      title = {Transferring climate change physical knowledge},
      journal = {Proceedings of the National Academy of Sciences},
      volume = {122},
      number = {15},
      pages = {e2413503122},
      year = {2025},
      doi = {10.1073/pnas.2413503122},
      URL = {https://www.pnas.org/doi/abs/10.1073/pnas.2413503122},
      eprint = {https://www.pnas.org/doi/pdf/10.1073/pnas.2413503122},
}

License

MIT License

Copyright (c) 2024 Francesco Immorlano

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.