Skip to content
Supplementary Materials for Tshitoyan et al. "Unsupervised word embeddings capture latent knowledge from materials science literature", Nature (2019).
Branch: master
Clone or download
vtshitoyan Merge pull request #5 from JohannesEbke/master
Fix typo 'bettery' -> 'battery'
Latest commit bc8774c Jul 11, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
mat2vec change do -> perform Jul 3, 2019
.gitignore first commit Jun 11, 2019
LICENSE Modify license copyright. Reorder Readme. Jun 23, 2019 fixing data copying bug Jun 20, 2019 Fix typo 'bettery' -> 'battery' Jul 10, 2019
requirements.txt Add version to requests. Jun 22, 2019 Add a script to download Word2vec model files from GCS. Jun 22, 2019

Supplementary Materials for "Unsupervised word embeddings capture latent knowledge from materials science literature", Nature (2019).

Tshitoyan, V., Dagdelen, J., Weston, L., Dunn, A., Rong, Z., Kononova, O., Persson, K. A., Ceder, G. and Jain, A.

doi: 10.1038/s41586-019-1335-8

Set up

  1. Make sure you have python3.6 and the pip module installed. We recommend using conda environments.
  2. Navigate to the root folder of this repository (the same folder that contains this README file) and run pip install -r requirements.txt. Note: If you are using a conda env and any packages fail to compile during this step, you may need to first install those packages separately with conda install package_name.
  3. Wait for all the requirements to be downloaded and installed.
  4. Run python install to install this module. This will also download the Word2vec model files. If the download fails, manually download the model, word embeddings and output embeddings and put them in mat2vec/training/models.
  5. Finalize your chemdataextractor installation by executing cde data download (You may need to restart your virtual environment for the cde command line interface to be found).
  6. You are ready to go!


Example python usage:

from mat2vec.processing import MaterialsTextProcessor
text_processor = MaterialsTextProcessor()
text_processor.process("LiCoO2 is a battery cathode material.")

(['CoLiO2', 'is', 'a', 'battery', 'cathode', 'material', '.'], [('LiCoO2', 'CoLiO2')])

For the various methods and options see the docstrings in the code.

Pretrained Embeddings

Load and query for similar words and phrases:

from gensim.models import Word2Vec
w2v_model = Word2Vec.load("mat2vec/training/models/pretrained_embeddings")

[('thermoelectrics', 0.8435688018798828), ('thermoelectric_properties', 0.8339033126831055), ('thermoelectric_power_generation', 0.7931368350982666), ('thermoelectric_figure_of_merit', 0.7916493415832 52), ('seebeck_coefficient', 0.7753845453262329), ('thermoelectric_generators', 0.7641351819038391), ('figure_of_merit_ZT', 0.7587921023368835), ('thermoelectricity', 0.7515754699707031), ('Bi2Te3', 0 .7480161190032959), ('thermoelectric_modules', 0.7434879541397095)]

Phrases can be queried with underscores:

w2v_model.wv.most_similar("band_gap", topn=5)

[('bandgap', 0.934801459312439), ('band_-_gap', 0.933477520942688), ('band_gaps', 0.8606899380683899), ('direct_band_gap', 0.8511275053024292), ('bandgaps', 0.818678617477417)]


# helium is to He as ___ is to Fe? 
    positive=["helium", "Fe"], 
    negative=["He"], topn=1)

[('iron', 0.7700884938240051)]

Material formulae need to be normalized before analogies:

# "GaAs" is not normalized
    positive=["cubic", "CdSe"], 
    negative=["GaAs"], topn=1)

KeyError: "word 'GaAs' not in vocabulary"

from mat2vec.processing import MaterialsTextProcessor
text_processor = MaterialsTextProcessor()
    positive=["cubic", text_processor.normalized_formula("CdSe")], 
    negative=[text_processor.normalized_formula("GaAs")], topn=1)

[('hexagonal', 0.6162797212600708)]

Keep in mind that words should also be processed before queries. Most of the time this is as simple as lowercasing, however, it is the safest to use the process() method of mat2vec.processing.MaterialsTextProcessor.


To run an example training, navigate to mat2vec/training/ and run

python --corpus=data/corpus_example --model_name=model_example

from the terminal. It should run an example training and save the files in models and tmp folders. It should take a few seconds since the example corpus has only 5 abstracts.

For more options, run

python --help


You can either report an issue on github or contact one of us directly. Try,, or

You can’t perform that action at this time.