Learning Piezoelectric Tensors through Strain-Conditioned Polarization Clusters.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Predicting tensorial properties like piezoelectricity is difficult due to coupled symmetry, rotational, and anisotropic constraints.
• Existing regression models struggle with these complex constraints, limiting their accuracy.

Purpose of the Study:

• To develop a novel physics-informed cluster graph neural network (PCGNN) for accurate piezoelectric tensor prediction.
• To explicitly encode the strain-polarization response pathway for improved physical grounding.

Main Methods:

• Developed a PCGNN that applies strain perturbations and reconstructs tensors via symmetry-consistent aggregation of local polarization clusters.
• Introduced a capsule transformer to identify symmetry-equivalent local environments and recover sparsity patterns.
• Trained and tested the model on the Materials Project dataset.

Main Results:

• PCGNN achieved a mean absolute error of 0.135 C/m² on the Materials Project test set.
• Outperformed existing models EATGNN (by 32.5%) and CGCNN (by 52.6%).
• Successfully identified candidate materials with strong piezoelectric responses.

Conclusions:

• The PCGNN provides a physically grounded approach for tensor learning and materials discovery.
• The model's ability to handle complex constraints opens new avenues for predicting material properties.
• Facilitates large-scale screening of unlabeled materials for piezoelectric applications.