Machine Learning a Simple Interpretable Short-Range Potential for Silica.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Accurate modeling of silica polymorphs is crucial for understanding material properties and processes.
• Existing models, including ab initio and force fields, face challenges in balancing accuracy and computational efficiency.
• Machine-learned potentials show promise but often retain high computational costs.

Purpose of the Study:

• To develop an accurate and computationally efficient model for silica polymorphs.
• To optimize parameters for a BKS-based Soules potential using reinforcement learning.
• To compare the performance of the new model against existing high-fidelity methods.

Main Methods:

• Utilized a reinforcement learning (RL) workflow to optimize an eight-dimensional parameter space for the Soules potential.
• Employed an experimental training dataset encompassing local and global structural features from 21 silica polymorphs.
• Compared the developed ML-Soules model with ML-BKS, GAP, and ab initio SCAN functional calculations.

Main Results:

• The ML-Soules model accurately predicts the relative energetic ordering and structural features of various silica polymorphs.
• Achieved significantly reduced computational expense compared to other high-quality models.
• Demonstrated reasonable accuracy in capturing the structure, density, and elastic constants of quartz and metastable polymorphs.

Conclusions:

• The ML-Soules model offers a promising balance of accuracy and computational efficiency for silica modeling.
• Reinforcement learning is effective for optimizing complex potential energy surfaces.
• Further enhancements to the Soules functional form could improve accuracy for both global and local features.