Critical assessment of machine-learned repulsive potentials for the density functional based tight-binding method: A...

Area of Science:

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

Background:

• Semi-empirical methods like density functional based tight-binding (DFTB) offer computational efficiency but are limited by simplified potentials.
• Developing transferable interatomic potentials is crucial for accurate simulations of materials properties.
• Machine learning (ML) presents a powerful framework for creating sophisticated atomic potentials.

Purpose of the Study:

• To investigate the feasibility of enhancing the semi-empirical DFTB method for pure silicon.
• To develop a general and transferable many-body repulsive potential using a common ML framework.
• To assess the performance of ML-driven potentials against traditional methods and fully ML potentials.

Main Methods:

• Utilizing atom-centered symmetry functions to describe atomic environments.
• Employing flexible neural networks to learn the many-body repulsive potential.
• Training the ML model on a diverse range of silicon systems, including bulk and clusters.
• Comparing results with existing DFTB parameterizations and fully machine-learned potentials.

Main Results:

• Achieved significant improvements in energetic, vibrational, and structural properties for bulk silicon systems.
• Encountered difficulties in accurately describing silicon clusters due to surface effects.
• Demonstrated the transferability of the developed ML-based potential across different silicon structures.
• Identified the impact of ML model scope on the accuracy of total energy calculations.

Conclusions:

• The developed many-body repulsive potential improves the semi-empirical DFTB method for bulk silicon.
• ML-based potentials show promise for transferable interatomic interactions but require careful consideration of system-specific effects like surfaces.
• This work highlights the potential and limitations of integrating ML into established computational materials science methods.