Physically Consistent Deep Learning Reveals Overestimation of Future Soil Drying

Transferring Physical Knowledge Reveals Overestimated Future Soil Drying This study uses Python 3.11.0 and TensorFlow 2.14.0. Folders Fig_1–Fig_4: Contain the data and scripts used to generate the corresponding figures in the manuscript. Folder model: Contains scripts for the transfer learning models (TLMs), including: unet.py: Defines and builds the TLMs. First_training.py: Pre-trains each TLM on 30 CMIP6 Earth System Models (ESMs) under three SSP scenarios (SSP245, SSP370, SSP585). Fine_tuning_ESMs.py: Fine-tunes each TLM using a perfect-model evaluation framework. Fine_tuning_observation.py: Fine-tunes each TLM using observational data. util.py: Defines variables and functions used in the scripts. Folder example_data: Provides example data for pre-training and fine-tuning using observational data. Data for fine-tuning using ESMs should be downloaded from the https://aims2.llnl.gov/search/cmip6/.