Work Function Database and ML Model

This repository contains the work function database (calculated by high-throughput density functional theory) and machine learning model to predict the work function of a surface, accompanying the preprint:

Discovery of materials with extreme work functions by high-throughput density functional theory and machine learning, Peter Schindler, Evan R. Antoniuk, Gowoon Cheon, Yanbing Zhu, and Evan J. Reed, arXiv:2011.10905

Database

The database csv files can easily be loaded using pandas.read_csv. The columns contain the following information:

• mpid: Materials Project ID
• miller: Miller index of surface hkl
• term: Termination number indexing (unique terminations are numbered starting at 1)
• surface_elements: List of atomic numbers present in the topmost atomic layer
• surface_elements_string: Same information as surface_elements but as a concatenated string of elements
• WF: The DFT-calculated work function in eV
• slab: The surface slab as a Pymatgen dictionary
• energy: Total energy of slab in eV
• Fermi: Fermi level in eV
• convergence: DFT convergence parameters, E_cutoff - k_x - k_y - k_z
• nsites: Number of atomic sites in the slab unit cell
• slab_thickness: Slab thickness in Angstroms
• nterm: Number of unique terminations for given orientation/material
• mirror: Whether or not there exists a mirror plane parallel to the surface

ML Model

The folder ML_model contains code to generate surfaces from bulk input structures, to featurize them, and to make work function predictions.

Requirements

The code has been tested with Python 3.6.6 and the following packages/versions: ase==3.20.1, pymatgen==2020.6.8, scikit-learn==0.23.1, pandas==1.0.1, joblib==0.16.0, but may run with newer versions as well.

How to use

To generate surfaces from the bulk and predict their work functions follow these steps:

1. Create input.csv file for your bulk input structure with two columns:
   1. material_id: The Materials Project ID or any other unique label for the bulk input structure
   2. cifs.conventional_standard: The conventional unit cell cif information (e.g. as pulled with the same column name tag using MPRester)
2. Run slab_generator.py which creates all unique surfaces/terminations up to a Miller index of 1 (then stored in file slabs-input.csv).
3. Run featurization_and_WF_predict.py to featurize the above created surfaces and predict the work function. The predicted work functions are stored in file slabs-input_predicted_WFs_0p3.csv under column WF_predicted.