ML Features 4 Materials and Molecules

This repo contains self-implemented classes for calculating ML features of materials and molecules. At the moment, only the following features are implemented:

• Revised Autocorrelation (RAC) features

Installation

To install the package, clone the repository and run the following command in the root directory:

pip install .

Usage

Revised Autocorrelation (RAC) features

RAC features are implemented based on the following paper: https://pubs.acs.org/doi/full/10.1021/acs.jpca.7b08750. Please credit the authors and this Github repo if you use this implementation.

RAC features map a molecular graph to a numerical vector. They are great if you need a numerical representation of a molecular graph for a machine learning model like a Random Forest. They work similar to a Graph Neural Network, but they don't train any weights and instead just sum all values, which makes them much simpler to use. They return a feature vector which can be used for example for scikit-learn machine learning models. Importantly, the current implementation does not support user-defined node properties or bond properties, but it would be possible to implement them if necessary.

The representation takes into account the connectivity (bonds) between the atoms and the atom types of each atom. The RAC features here use multiple atomic properties such as electronegativity, atomic number, and atomic mass to calculate the autocorrelation of a graph. The RAC features are calculated up to a certain depth (e.g. 4 as a good default), which is a hyperparameter that can be set by the user. The RAC features are calculated for each depth and for each atomic property. In order to normalize the RAC features so that they have the same length no matter the number of nodes in the graph, 4 different statistical measures are applied to each atomic property, sum, std, min and max. The original implementation in the paper above uses only the sum, this was generalized here. The RAC features are calculated for the following atomic properties:

• electronegativity
• row
• group
• atomic_mass
• electron_affinity
• min_oxidation_state
• max_oxidation_state
• ionization_energy
• nuclear_charge
• ident
• topology
• size

The following code snippet demonstrates how to calculate RAC features for a CH molecule:

    import networkx as nx
    from MLFeatures4Materials.features.rac import RAC
    # Example graph:
    G = nx.Graph()
    G.add_nodes_from([(0, {'node_label': 'C'}),
                      (1, {'node_label': 'H'})])
    G.add_edges_from([(0, 1)])
    features, labels = RAC(depth=4).molecule_autocorrelation(graph=G, return_labels=True, element_label='node_label')
    print('Features:', features)
    print('Labels:', labels)

The output of the code snippet above is:

  Features: [11.342500000000001, 11.22, 0.0, 0.0, 0.0, 0.8312499999999994, 0.0, ...]
  Labels: ['chi-0-sum', 'chi-1-sum', 'chi-2-sum', 'chi-3-sum', 'chi-4-sum', 'chi-0-std', 'chi-1-std', ...]

License

This project is licensed under the MIT License - see the LICENSE file for details. This means that you can use this code for free for any purpose, but the authors of this package cannot be hold responsible for anything.