How close are the classical two-body potentials to ab initio calculations? Insights from linear machine learning...

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Condensed Matter Theory

Background:

• Classical force fields are crucial for simulating materials but often lack accuracy.
• Accurate atomic interactions are essential for predicting material properties.

Purpose of the Study:

• To develop a machine learning approach for extracting accurate pair atomic interactions from ab initio calculations.
• To create a more reliable classical force field for amorphous silica.

Main Methods:

• A linear machine learning force matching approach was employed.
• Local feature representation was used to define potentials as functions of interatomic distances.
• The method was validated using ab initio calculations on amorphous silica.

Main Results:

• The derived force field potentials for Si-Si, Si-O, and O-O interactions differ significantly from existing potentials.
• Simulations using the new force field yield a lower glass transition temperature (Tg ~ 1800 K) and positive liquid thermal expansion.
• Existing classical force fields may produce artifacts in simulated silica properties, such as unusually high Tg and negative thermal expansion.

Conclusions:

• The proposed method provides a fundamental way to evaluate two-body potentials against ab initio data.
• This approach efficiently guides the development of accurate classical force fields for amorphous materials.
• Classical force fields commonly used for silica may not accurately represent atomic interactions, potentially leading to artifacts in simulated properties.