Silicon Liquid Structure and Crystal Nucleation from Ab Initio Deep Metadynamics.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Empirical potentials struggle to accurately model silicon's phase transitions between semiconducting solid and metallic liquid states.
• Nucleation events in silicon crystallization occur on timescales inaccessible to standard ab initio molecular dynamics simulations.

Purpose of the Study:

• To develop a computationally efficient and accurate method for studying silicon crystallization and nucleation.
• To improve the modeling of silicon's complex phase diagram and thermodynamic properties.

Main Methods:

• Training a deep neural network potential using metadynamics simulations with a classical potential.
• Introducing a novel collective variable based on the Debye structure factor to describe nucleation dynamics.
• Utilizing the strongly constrained and appropriately normed (SCAN) exchange-correlation functional for accurate energy calculations.

Main Results:

• The deep neural network potential accurately reproduces silicon's properties and enables efficient simulation of crystallization.
• The new collective variable effectively captures long-range order information for nucleation dynamics.
• Calculated thermodynamic properties near the melting point show good agreement with experimental data.

Conclusions:

• The developed deep neural network potential offers a powerful tool for studying silicon crystallization.
• The study provides new insights into the early stages and mechanisms of silicon nucleation.
• This approach bridges the gap between classical and ab initio methods for materials simulation.