Virtual node graph neural network for full phonon prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Understanding material structure-property relationships is key for designing new materials.
• Machine learning (ML) has advanced this field, but faces challenges in model generalizability and predicting properties with variable output dimensions.

Purpose of the Study:

• To address challenges in ML-based material property prediction.
• To introduce a novel virtual node graph neural network (VNGNN) for predicting phonon properties.
• To enable efficient and accurate prediction of phonon spectra and band structures.

Main Methods:

• Developed three virtual node approaches within a graph neural network framework.
• Applied the VNGNN to predict Gamma-phonon (Γ-phonon) spectra and full phonon dispersion from atomic coordinates.
• Compared the VNGNN approach against machine-learning interatomic potentials (MLIPs).

Main Results:

• Achieved orders-of-magnitude higher efficiency than MLIPs with comparable or better accuracy.
• Generated a database of Γ-phonon spectra for over 146,000 materials.
• Successfully predicted phonon band structures for zeolites.

Conclusions:

• The virtual node method offers a flexible and generic approach for ML-driven materials design.
• Enables rapid, high-quality prediction of phonon band structures for designing materials with specific phonon properties.
• Advances the application of graph neural networks in materials science for property prediction.