Machine learning interatomic potential for silicon-nitride (Si3N4) by active learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Silicon nitride (Si3N4) is crucial in automotive, aerospace, and semiconductor industries.
• Reliable interatomic potentials for atomistic simulations of Si3N4, especially its amorphous phase, are scarce.

Purpose of the Study:

• To develop an accurate and efficient machine learning interatomic potential for silicon nitride.
• To enable atomistic-scale studies of amorphous silicon nitride properties.

Main Methods:

• Employed Gaussian Approximation Potential (GAP) method with an active learning strategy.
• Generated initial dataset using an empirical potential, refined with Density Functional Theory (DFT) calculations.
• Utilized an iterative re-training algorithm for on-the-fly learning and potential improvement.

Main Results:

• Achieved a mean absolute error of 8 meV/atom compared to DFT for liquid and amorphous structures.
• Enabled molecular dynamics simulations 3-4 orders of magnitude faster than DFT.
• Demonstrated excellent agreement with experimental results for silicon nitride.

Conclusions:

• The developed machine learning potential offers a significant advancement for atomistic simulations of silicon nitride.
• The potential provides a computationally efficient and accurate tool for studying amorphous silicon nitride.
• The publicly available potential facilitates further research in materials science and engineering.