Representing Born effective charges with equivariant graph convolutional neural networks.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Graph convolutional neural networks (GCNs) are powerful tools for predicting material properties.
• Representing tensorial properties requires network weights and descriptors to be invariant to reference frame changes.
• Existing GCNs may not inherently satisfy these transformation rules for tensorial properties.

Purpose of the Study:

• To develop a physics-informed, equivariant graph convolutional neural network (EGCNN).
• To explicitly encode rotational symmetry for accurate prediction of tensorial material properties.
• To demonstrate the network's performance on predicting atomic Born effective charges.

Main Methods:

• Developed an equivariant graph convolutional neural network architecture.
• Incorporated equivariant weights and descriptors to respect crystal rotational symmetries.
• Applied the EGCNN to predict atomic Born effective charges in various materials.

Main Results:

• The EGCNN successfully respects rotational symmetries of crystal structures.
• The network provides accurate tensorial outputs for target material properties.
• Demonstrated good performance and generalization ability on perovskite oxides, Li3PO4, and ZrO2.

Conclusions:

• Equivariant graph convolutional neural networks are effective for learning tensorial material properties.
• The proposed EGCNN ensures physical constraints, leading to reliable predictions.
• This approach enhances the accuracy and generalizability of machine learning models in materials science.