zjukeliu / atom2vec

A python implement of Atom2Vec: a simple way to describe atoms for machine learning

PyAtom2Vec

A python implement of Atom2Vec: a simple way to describe atoms for machine learning

Background

Atom2Vec is first proposed on Zhou Q, Tang P, Liu S, et al. Learning atoms for materials discovery[J]. Proceedings of the National Academy of Sciences, 2018, 115(28): E6411-E6417.

It is a powerful but simple method to transfer atoms into vectors, quite similar to Word2Vec in NLP.

Requirements

To run this program, you will need Scipy and Numpy packages. If you want to generate your own dataset, you may also need Requests package for web requests.

• Anaconda environment is highly recommended.

If you have installed pip, you may use these commands to install these packages.

# on Linux
pip3 install scipy numpy requests
# on Windows
pip install scipy numpy requests

How To Use

from Atom2Vec import Atom2Vec

# data_file: path to the dataset file
# vec_length: length of atom vector you want
atoms_vec = Atom2Vec(data_file, vec_length)
atoms_vec.saveAll()

Output:

Generating index 77402/77402 -- Complete!
Building matrix  -- Complete!
SVD -- Complete!

Also, this package contains a dataset, which was obtained from Material Project using GetMP.py. The raw response is stored in string_2.json and string_3.json. Then the response is further processed by Preprocess.py, whose result is saved to string.json for further use.

Output File

The output is kept in atoms_vec.txt and atoms_index.txt.

• atoms_vec.txt contains a M * N matrix. M is the index of the atoms. N is the length of atom vector. Each row represents a vector describing certain atom.

• atoms_index.txt contains a M * 1 matrix. Each row contains a integer, which is the atomic number of certain atom. It tells which atom each row represents in atoms_vec.txt.

Test Program

Atom2Vec.py also contains a simple test, which can be run by

# Linux
python3 Atom2Vec
# Windows
python Atom2Vec

If the program can run normally, it will exit with no errors raised.

Interactive Similarity Map

We can calculate cosine distance to quantify similarity between every atom.

You can find a interactive similarity map on https://www.yuxingfei.com/src/similarity.html