BNM-CDGNN: Batch Normalization Multilayer Perceptron Crystal Distance Graph Neural Network for Excellent-Performance...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Graph neural networks (GNNs) excel at predicting crystal properties by capturing structural features.
• Existing GNN models struggle with performance limitations in hidden layers post-pooling.
• Accurate crystal property prediction is crucial for materials discovery and development.

Purpose of the Study:

• To introduce a novel GNN model, BNM-CDGNN, designed to overcome limitations in existing GNN architectures for crystal property prediction.
• To enhance the model's ability to learn precise geometric information and improve training performance.

Main Methods:

• Proposed the batch normalization multilayer perceptron crystal distance graph neural network (BNM-CDGNN).
• Encoded crystal geometry using atomic distance vectors.
• Employed radial basis function as an attention mask in graph convolutional layers for rotation invariance.
• Integrated average pooling and batch normalization in multiple hidden layers.

Main Results:

• BNM-CDGNN demonstrated significantly enhanced accuracy in crystal property prediction.
• The model effectively captured geometric information and rotation invariance.
• Improved handling of hidden layers after pooling contributed to better training performance.

Conclusions:

• BNM-CDGNN offers a superior approach to crystal property prediction compared to established models like SchNet and MPNN.
• The novel architecture effectively leverages geometric features and improves model training.
• This advancement holds promise for accelerating materials science research through accurate property prediction.