SLI-GNN: A Self-Learning-Input Graph Neural Network for Predicting Crystal and Molecular Properties

Area of Science:

• Computational Materials Science
• Machine Learning for Materials Discovery
• Graph Neural Networks (GNNs)

Background:

• Crystalline and molecular structures are non-Euclidean data, posing challenges for traditional modeling.
• Graph neural networks (GNNs) offer a powerful approach for representing and analyzing materials data.
• Accelerating the discovery of new materials requires efficient and accurate predictive models.

Purpose of the Study:

• To introduce a novel self-learning-input GNN (SLI-GNN) framework for unified property prediction in crystals and molecules.
• To enhance the efficiency and accuracy of material property prediction using GNNs.
• To accelerate the discovery of new materials through advanced computational methods.

Main Methods:

• Development of a self-learning-input GNN (SLI-GNN) framework.
• Implementation of a dynamic embedding layer for self-updating input features during neural network iteration.
• Integration of the Infomax mechanism to maximize mutual information between local and global features.

Main Results:

• The SLI-GNN framework achieves ideal prediction accuracy with reduced input requirements and optimized message passing neural network (MPNN) layers.
• Model evaluations on the Materials Project and QM9 datasets demonstrate performance comparable to existing GNNs.
• The framework shows excellent performance in predicting material properties.

Conclusions:

• The proposed SLI-GNN framework offers a promising approach for accelerating new material discovery.
• SLI-GNN provides an efficient and accurate method for predicting properties of both crystalline and molecular materials.
• The dynamic embedding and Infomax mechanisms contribute to the framework's enhanced predictive capabilities.