Realistic Atomistic Structure of Amorphous Silicon from Machine-Learning-Driven Molecular Dynamics.

Area of Science:

• Materials Science
• Computational Materials Science
• Condensed Matter Physics

Background:

• Amorphous silicon (a-Si) is a crucial noncrystalline material with applications in electronics.
• Despite extensive study, its precise atomistic structure remains poorly understood.
• Accurate structural models are essential for predicting and optimizing a-Si properties.

Purpose of the Study:

• To develop and validate a machine-learning-based interatomic potential for modeling amorphous silicon.
• To generate high-quality structural models of a-Si with minimal defects.
• To achieve unprecedented agreement between simulated and experimental data for a-Si.

Main Methods:

• Utilized a machine-learning interatomic potential for atomistic simulations.
• Employed simulated cooling from the melt at a controlled rate (10^11 K/s, 10 ns timescale).
• Generated large-scale (4096-atom) a-Si models for structural analysis.

Main Results:

• Developed an a-Si model with less than 2% defects.
• Achieved excellent agreement with experimental excess energies, diffraction data, and 29Si NMR chemical shifts.
• Successfully reproduced the first sharp diffraction peak (FSDP) magnitude in the structure factor.

Conclusions:

• Machine-learning potentials are highly effective for accurate amorphous silicon structure modeling.
• Controlled cooling rates are critical for obtaining high-quality a-Si models.
• This approach advances the understanding of amorphous materials and their properties.