pGNN : Proxy Graph Neural Network Featurizer

pGNN is a featurizer tool that leverages pretrained graph-neural networks (GNNs) as a proxy to extract valuable chemical information from materials structures. It aims to enhance the performance of materials property prediction models by generating highly meaningful features that can be used in a variety of machine learning tasks.

pGNN offers multiple types of feature extraction tools to provide flexibility for various use cases:

• MVLFeaturizer: The MVLFeaturizer provides features extracted from various pretrained MEGNet models. These models have been trained on diverse datasets and offer reliable features for a wide range of properties. Users can choose between different intermediate layers to extract features from the regression head of the MEGNet models with either a 16-neuron or 32-neuron configuration.

• CustomMEGNetFeaturizer: The CustomMEGNetFeaturizer allows users to load custom models from the custom_models/ directory. These models can incorporate encoded chemical information derived from extensive featurizers such as MatMiner (l-MM, MatMinerEncoded\_v1 model) and Orbital Field Matrix (l-OFM, OFMEncoded\_v1 model), providing tailored feature extraction for diverse use cases.

• AdjacentMEGNetFeaturizer: The AdjacentMEGNetFeaturizer is designed to train a MEGNet model on the fly for the user's dataset. This is particularly useful for ensuring that the features extracted are closely aligned with the dataset at hand, thus reducing potential bias and improving model accuracy. Similar to the other featurizers, users can choose to extract features from either the 16-neuron or 32-neuron layer.

To install pGNN, clone the repository and use the following command:

pip install -r requirements.txt

Ensure you have all the necessary dependencies installed, including TensorFlow, scikit-learn, and the necessary MEGNet libraries.

Here's an example of how to use pGNN to extract features from a list of structures:

from pgnn import MVLFeaturizer, AdjacentMEGNetFeaturizer
from pgnn import get_Custom_MEGNetFeatures

# Initialize featurizers
mvl32 = MVLFeaturizer(layer_name='layer32')
mvl16 = MVLFeaturizer(layer_name='layer16')
adj_megnet = AdjacentMEGNetFeaturizer(layer_name='layer32')

# Example structures
structures = pd.DataFrame({'structure': [...]})  # Replace '...' with your actual list of structures

# Extract features using MVLFeaturizer
features_32 = mvl32.get_features(structures)
features_16 = mvl16.get_features(structures)

# Train the AdjacentMEGNetFeaturizer on the fly
# Provide in 'targets' some property values corresponding to these structures
adj_megnet.train_adjacent_megnet(structures, targets=[...])

# Extract features using the trained AdjacentMEGNetFeaturizer
features_adj = adj_megnet.get_features(structures)

# Load features using custom model
features = get_Custom_MEGNetFeatures(structures, model_type='OFMEncoded_v1')

We welcome contributions to improve the pGNN tool, including adding more pretrained models or enhancing the featurization techniques. Please feel free to submit pull requests or create issues for discussion.

This project is licensed under the MIT License.

The pGNN tool is built on top of the MEGNet framework, leveraging its pretrained models to provide state-of-the-art features for materials science research. We thank the Materials Virtual Lab for making the pretrained models publicly available.