Accelerated discovery of stable compositions in inorganic materials with Bayesian optimisation of the chemical phase fields

3.02.2021 Andrij Vasylenko

Functionality

The modes of running the code are:

1. mode = 'path' can calculate a 'would-be' Bayesian Optimisation path for the previously calculated compositions in the phase field

2. mode = 'suggest' suggests new compositions for calculations, based on the precomputed results.

3. mode = 'generate' generates the list of candidate compositions that can be edited and further used for limiting the candidates to the particular compositions only in a subsequent run in 'suggest' mode

Requirements

python-3.9

Dependencies:

numpy; pandas; pymatgen; scikit-learn GpyOpt

Dependencies can be installed automatically during installation.

Installation

pip install .

Usage

1. Prepare a 2-column table, where each row has a composition and its value of Total Energy as a .csv file. Make sure you include reference compositions in the phase field.
2. Modify input_config.yaml file accordingly, providing names of the file, atoms, and their oxidation states.
3. run:

python -m phasebo

will use input_config.yaml by default

or

python -m phasebo --config path/to/my_config.yaml

Example

The default run with input_config.yaml results int the outputs in example

Reference

Please consider citing this tool: A. Vasylenko et al, Inferring energy–composition relationships with Bayesian optimization enhances exploration of inorganic materials. The Journal of Chemical Physics 160, 5, 054110 (2024)

The code is based on GPyOpt implementation of Bayesian optimisation algorithm. @Misc{gpyopt2016, author = {The GPyOpt authors}, title = {GPyOpt: A Bayesian Optimization framework in Python}, howpublished = {\url{http://github.com/SheffieldML/GPyOpt}}, year = {2016} }

Parameters of the input configuration file

parameter	value
mode	(default: 'suggest') Mode of calculations: the best path so far ('path'); suggest next compositions for CSP based on the available results ('suggest'); generate candidate compositions into candidates_list.csv ('generate')
inputfile	(default: LiSnSCl_700eV.csv) Input file. A table of compositions and their total energies.
compositionfile	(default: None) Input file. A list of candidate compositions (formulas) to consider. If not provided, the candidates will be generated automatically.
excludefile	(default: None) Input file. A list of compostions (formulas) to exclude from convex hull calculations as well as from candidates. If not provided, no candidates are excluded.
ions	(default: {'Li':1,'Sn':4,'S':-2,'Cl':-1}) Ions and oxidation states.
seeds_type	(default: 'random') Method to choose seeds in mode == 'path': 'segmented': Seeds are picked from a segmented phase field. 'random' seeds are selected randomly.
disect	(default: 4) Number of sections of the phase field (disect x disect), from which the 'segmented' seeds are selected.
N_atom	(default: 24) Maximum number of atoms per unit cell in suggested compositions (in 'suggest' and 'generate' modes)
max_iter	(default: 10) Maximum number of iterations.