Tool to quickly create a composition-based feature vectors from materials datafiles.
The source code is currently hosted on GitHub at: https://github.com/kaaiian/CBFV
Binary installers for the latest released version are available at the Python Package Index (PyPI)
# PyPI
pip install CBFV
The CBFV package assumes your data is stored in a pandas dataframe of the following structure:
| formula | target |
|---|---|
| Tc1V1 | 248.539 |
| Cu1Dy1 | 66.8444 |
| Cd3N2 | 91.5034 |
To featurize this data, the generate_features function can be called as follows:
from CBFV import composition
X, y, formulae, skipped = composition.generate_features(df)
The featurization scheme can be adjusted by calling the the elem_prop parameter. The following featurization schemes are included within CBFV:
- jarvis
- magpie
- mat2vec
- oliynyk (default)
- onehot
- random_200
Duplicate formula handeling is controlled by the drop_duplicates parameter. It is set to False by default to preserve datapoints containing variation outside of their formula. For example, heat capacity measurements performed for the same material at different temperatures.
The extend_features parameter is used to specify whether columns outside of ['formula', 'target'] should be considered during featurization. It is set to False by default to exclude nuisance information from consideration. Setting extend_features=True would allow additional information (i.e. ['temperature', 'pressure']) to be preserved.
The sum_feat parameter specifies whether to calculate the sum features when generating the CBFVs for the chemical formulae. It is set to False by default.
Calling generate_features with these parameters can be implemented as follows:
| formula | target | temp |
|---|---|---|
| Tc1V1 | 248.539 | 373 |
| Tc1V1 | 66.8444 | 473 |
| Cd3N2 | 91.5034 | 273 |
from CBFV import composition
X, y, formulae, skipped = composition.generate_features(df,
elem_prop='magpie',
drop_duplicates=False,
extend_features=True,
sum_feat=True)