Proxy-informed climate model emulators for improved particle filter assimilation frameworks

Version 1 for publication "Proxy-informed climate model emulators for improved particle filter assimilation frameworks". User guide: This repository contains the code, processed climate-model datasets, trained models, and precomputed results used to reproduce the analyses and figures presented in the publication. The repository is organised into two main archives: Code.zip and Data.zip. Both archives should be extracted into the same parent directory, producing two folders: Codes/ and DATA/. Code.zip The Codes/ directory follows the main steps of the methodological framework: Dimension reduction/ contains the implementations of Empirical Orthogonal Functions (EOFs) and autoencoders (AEs), used to represent high-dimensional climate fields in reduced spaces. Emulators/ contains the notebooks used to train the Linear Inverse Model (LIM), the Autoencoder-based Recurrent Neural Network (AERNN), and the Proxy-informed AERNN (PAERNN). Particle Filters/ contains the Sequential Importance Resampling (SIR) particle-filter implementations using the LIM, AERNN, and PAERNN emulators, together with a reference SIR-0 experiment and the shared particle-filter functions. Cross-model experiments/ contains the notebooks for historical CMIP6 pretraining, multimodel fine-tuning, and particle-filter experiments evaluating the transferability of emulators trained on IPSL-CM6A-LR and on multiple climate models. Figures/ contains the notebooks used to reproduce Figures 2–11 and Supplementary Figure A1 from the saved experimental outputs. Figures 4–6 are generated using a common notebook. The Codes/ directory also includes article_paths.py, which manages the paths to the datasets and experiment outputs, and a README.md providing additional instructions. Data.zip The DATA/ directory contains the datasets, trained models, and precomputed outputs required for the analyses. It is organised into three main folders: Inputs/ contains the processed IPSL-CM6A-LR past2k ensemble (10 simulations of 1,300 years), the historical CMIP6 simulations used for emulator pretraining, four past1000 climate-model simulations used for multimodel fine-tuning, the MIROC-ES2L simulation used for cross-model evaluation, the proxy network, and volcanic forcing data. Models/ contains the trained autoencoders, recurrent neural networks, LIM parameters, EOF solvers, normalisation parameters, and observational-error estimates used in the experiments. Outputs/ contains the precomputed particle-filter reconstructions, climate indices, evaluation metrics, and sensitivity results required to reproduce the figures and tables presented in the publication. Reproducing the results To reproduce the published figures, the recommended starting point is the Figures/ directory. The notebooks directly load the precomputed results from DATA/Outputs/, avoiding the need to repeat computationally expensive emulator training and particle-filter assimilation experiments. To reproduce the methodological workflow, dimensionality reduction can first be performed using the notebooks in Dimension reduction/. The reduced representations are then used to train climate emulators with the notebooks in Emulators/. These emulators can subsequently be used for sequential paleoclimate data assimilation through the scripts in Particle Filters/. The Cross-model experiments/ directory provides the additional training and evaluation procedures used to investigate the generalisation of these methods across climate models. When Codes/ and DATA/ are extracted into the same parent directory, the scripts automatically locate the required datasets. New experimental outputs are stored separately from the precomputed results supplied in DATA/. The figure-reproduction notebooks use existing outputs, whereas reproducing the full training and assimilation experiments requires additional computational resources. The requirements.txt file lists the main Python dependencies. The underlying climate-model simulations are available through the CMIP6 archive and ESGF data portals.