DynStabNet: A Deep Learning Framework for Fast Dynamical Stability Prediction of Crystal Structures

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Semiconductor materials are crucial for electronic, optoelectronic, and energy applications.
• Density Functional Theory (DFT)-based phonon calculations accurately assess structural dynamical stability but are computationally expensive for large-scale screening.
• Accelerating materials discovery requires faster methods for evaluating structural stability.

Purpose of the Study:

• To develop DynStabNet, an E(3)-equivariant graph neural network (E3GNN) framework for rapid prediction of material dynamical stability.
• To create a fast surrogate model that bypasses computationally intensive phonon calculations during inference.
• To significantly accelerate the materials design pipeline by enabling early elimination of unstable structures.

Main Methods:

• Developed DynStabNet, an E(3)-equivariant graph neural network (E3GNN) trained on phonon-informed data.
• Utilized a crystal structure generation model to create diverse candidate structures.
• Employed a pretrained machine learning potential to rapidly compute phonon spectra for training data generation.

Main Results:

• DynStabNet achieves 97% accuracy in predicting dynamical stability.
• The model reduces structure evaluation time from hours to approximately 1 millisecond.
• Demonstrated the capability to rapidly eliminate unstable configurations early in the materials design process.

Conclusions:

• DynStabNet offers an efficient framework for large-scale materials screening.
• The E(3)-equivariant architecture effectively captures structure-stability relationships.
• This approach significantly accelerates the discovery of stable semiconductor materials.