ASUGNN

Asymmetric Unit-Based Graph Neural Network for Crystal Property Prediction

---

Overview

ASUGNN is a Graph Neural Network (GNN) designed for crystal property prediction, leveraging asymmetric unit-based graph representations. This approach efficiently captures the structural and chemical properties of crystals, making it highly suitable for materials informatics tasks.

• HMdataset.db: A zero-shot test dataset You can test your model's zero-shot capability by predicting the formation energy per atom using the HMdataset (1,100 data) and upload your results on Kaggle : https://www.kaggle.com/competitions/asugnn.
• Source Code : src
• Tutorial : notebook

Graph Embedding of ASUGNN

The graph embedding is performed on each entry in the database using the Crylearn package. It extracts structural information about the crystal, including node embeddings, the ASU matrix, a distance matrix, and the simulated PXRD pattern.

Requirements

• Python 3.9.19
• Crylearn
• ase

Example Usage

Here is an example of how to use the Crylearn package to extract graph embedding data from a crystal database:

from Crylearn import cry2graph
from ase.db import connect

# Connect to the demo database
database = connect('demo.db')
entry_id = 1

# Parse the entry and extract the graph embedding
N, ASUAM, DAM, PXRD = cry2graph.parser(database, entry_id).get()

Output

• N (np.ndarray): The node embeddings, where each node (atom) is represented by a 106-dimensional feature vector (shape: N x 106).
• ASUAM (np.ndarray): The Asymmetric Unit Adjacency Matrix (shape: N x N).
• DAM (np.ndarray): The distance matrix between nodes in Cartesian coordinates (shape: N x N).
• PXRD (np.ndarray): The simulated diffraction pattern of the crystal, capturing global structural information (shape: 140-dimensional).

Explanation

• The node embeddings represent atomic features within the lattice.
• The distance matrix provides the distances between pairs of atoms.
• The PXRD vector represents the simulated diffraction pattern, offering a condensed representation of the crystal structure.

---

License

This project is licensed under the MIT License - see the LICENSE file for details.