Accurate and Transferable Machine Learning Potential for Molecular Dynamics Simulation of Sodium Silicate Glasses

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Simulating glass behavior requires accurate interatomic potentials.
• Existing empirical potentials struggle with broad compositional and temperature ranges.
• Machine learning offers a promising avenue for developing more robust potentials.

Purpose of the Study:

• To develop an accurate and transferable machine learning potential for binary sodium silicate glasses.
• To enable simulations across a wide range of compositions (0-50% Na2O) and temperatures (300-3000 K).
• To outperform existing empirical potentials in predicting glass structures and properties.

Main Methods:

• Utilized a neural network algorithm (DeePMD code) to approximate the potential energy surface based on local atomic geometry.
• Trained the model on a large dataset of total energies and atomic forces from density functional theory (DFT) calculations.
• Employed classical molecular dynamics (MD) simulations to generate training structures at various temperatures.

Main Results:

• Developed a robust and transferable machine learning potential for sodium silicate glasses.
• The ML potential accurately reproduces structures and properties like bond angle distribution, total distribution functions, and vibrational density of states.
• Outperformed empirical potentials in predicting key glass characteristics.

Conclusions:

• The developed machine learning potential provides a significant advancement for simulating sodium silicate glasses.
• The approach demonstrates high accuracy and transferability across compositional and temperature ranges.
• This methodology paves the way for simulating complex multicomponent oxide glasses with near ab initio accuracy at reduced computational cost.