Predicting Structural Properties of Pure Silica Zeolites Using Deep Neural Network Potentials

Area of Science:

• Materials Science
• Computational Chemistry
• Nanotechnology

Background:

• Machine learning potentials (MLPs) are revolutionizing multiscale atomistic modeling by accurately describing complex potential energy surfaces (PESs).
• Accurate modeling of zeolite dynamics is crucial for understanding and designing nanoporous materials.

Purpose of the Study:

• To train a high-accuracy MLP for modeling silica framework dynamics using a comprehensive dataset.
• To evaluate the MLP's performance in predicting various material properties.

Main Methods:

• Trained a DeePMD-kit MLP using an extensive density functional theory (DFT) dataset (Si-ZEO22) comprising 350,000 calculations across 219 zeolite topologies.
• Evaluated MLP performance by calculating structural properties, energy-volume trends, phonon density of states, and stress-strain relationships.

Main Results:

• The MLP demonstrated impressive agreement with DFT for predicting structural properties, energy-volume trends, and phonon density of states.
• The model achieved reasonable predictions for stress-strain relationships, even without explicit DFT stress data during training.
• Highlighted the MLP's capability to capture the inherent flexibility of zeolite frameworks.

Conclusions:

• MLPs can achieve near-ab initio accuracy for modeling nanoporous materials like zeolites.
• The developed MLP is a powerful tool for advancing research in zeolite dynamics and materials design.
• Motivates further development of MLPs for complex nanoporous materials.