README.md

DPOSE

Contains supplementary code and data for the
"Uncertainty quantification by direct propagation of shallow ensembles"

• manuscript: http://arxiv.org/abs/2402.16621

The repository is organized as follows:

• ./UCI_experiments/ contains the regression experiments on UCI datasets
• ./Atomistic_experiments/ contains the regression experiments on atomistic data
• ./Data/ contains additional DFT calculations that were generated for this work
• ./Plots_and_analysis/ contains the figures and analysis notebooks for this work

UCI_experiments

This part of the repository is structures as follows:

• data: contains the data used in the experiments, for setup see the README.md in the directory
  and run the load.pyscript to download the UCI data
• models: contains the source code for the models and training routines
• experiments: contains the individual experiments using different architectures on the UCI datasets

Data

This part of the repository is structures as follows:

• Surfaces_DFT: contains the DFT calculations of liquid water surfaces
• CP2K_inputs: contains the input files for the CP2K calculations
• H2O_size_extensive_data: contains the data for the experiments on MD snapshots of liquid water of increasing size

Atomistic_experiments

This part of the repository is structures as follows:

• data: contains the data used in the experiments, run the load.py scripts in each subdirectory, to:
  • download the data, from various source (ie MaterialsCloud or github)
  • obtain train test splits
  • preprocess the data (only needed for liquid water data)
• materials_model_predictions: contains the predictions of the different models
• trajectories_H2O.zip: Unzip this file to obtain the trajectory data for the liquid water experiments

Plots_and_analysis

This part of the repository is structures as follows:

• figure 2-8: Each subdirectory contains the notebooks to generate the respective figures from the paper
• si_plot 1-3: Each subdirectory contains the notebooks to generate the respective supplementary figures from the paper
• table_materials: Contains the notebook to generate table II from the paper
• heat_capacity: Contains the notebook to generate the heat capacity data from the paper

Convergence_toy

This part of the repository is structures as follows: Contains experiments of the convergence of MVE and shallow ensembles on a challenging toy problem