[Paper Review] Training a Machine-Learning Driven Gaussian Approximation Potential for Si-H Interactions

Davis Unruh, Reza Vatan Meidanshahi+2 more|arXiv (Cornell University)|Jun 5, 2021

Thin-Film Transistor Technologies4 citations

TL;DR

This paper presents a machine-learning-driven Gaussian approximation potential (GAP) for Si-H interactions, trained on density functional theory (DFT) data for energies, forces, and stresses. The potential achieves DFT-level accuracy at a fraction of the computational cost, enabling high-fidelity simulations of hydrogenated amorphous and liquid silicon with excellent agreement to DFT and experimental structural data.

ABSTRACT

Hydrogenation of amorphous silicon is critical to reducing defect densities, passivating mid-gap states and improving photoconductivity. However, achieving high accuracy with low computational cost in the treatment of Si-H interactions in atomistic simulations has been historically challenging. Here we introduce a machine-learning driven Gaussian approximation potential for atomistic simulations of various hydrogenated phases of silicon. Trained on density functional theory (DFT) measurements of energies, forces and stresses, this potential enables materials simulations of hydrogenated silicon with DFT-level accuracy but significantly reduced computational expense. We demonstrate the capabilities of the potential by using it to create hydrogenated liquid and amorphous silicon, and validating the structural measurements with excellent agreement against those of atomic configurations produced by density functional theory and experiment. These validations highlight the promise of using the potential for realistic and accurate simulations of a variety of hydrogenated silicon structures, particularly bulk a-Si:H and c-Si/a-Si:H heterojunctions.

Motivation & Objective

• To address the long-standing challenge of achieving high accuracy with low computational cost in modeling Si-H interactions in atomistic simulations.
• To develop a machine-learning potential that accurately captures the complex bonding and structural behavior of hydrogenated silicon across various phases.
• To enable reliable simulations of amorphous silicon hydride (a-Si:H) and c-Si/a-Si:H heterojunctions for materials design and characterization.
• To validate the potential against both DFT and experimental structural measurements to ensure physical realism and predictive power.

Proposed method

• The potential is constructed using a Gaussian process regression framework trained on a dataset of DFT-calculated energies, forces, and stresses for diverse Si-H configurations.
• Atomic environments are described using many-body atomic descriptors, enabling accurate representation of local chemical environments in silicon hydrides.
• The model is optimized to minimize prediction errors for energy, forces, and stress, ensuring consistency with quantum mechanical reference data.
• The trained potential is used to perform large-scale molecular dynamics simulations of hydrogenated liquid and amorphous silicon.
• Structural properties such as radial distribution functions and coordination numbers are computed and compared to DFT and experimental results.
• The framework allows for systematic evaluation of the potential's transferability across different phases and hydrogen concentrations.

Experimental results

Research questions

• RQ1 Can a machine-learning potential trained on DFT data accurately reproduce the structural and energetic properties of hydrogenated silicon across amorphous, liquid, and crystalline phases?
• RQ2 To what extent does the GAP model maintain DFT-level accuracy while reducing computational cost in Si-H systems?
• RQ3 How well does the potential predict structural features such as bond lengths, coordination numbers, and radial distribution functions compared to DFT and experiment?
• RQ4 Can the potential reliably simulate complex heterostructures like c-Si/a-Si:H interfaces with physical fidelity?

Key findings

• The machine-learning GAP potential achieves DFT-level accuracy in predicting energies, forces, and stresses for Si-H systems.
• Structural properties such as radial distribution functions and coordination numbers from the GAP simulations show excellent agreement with both DFT and experimental measurements.
• The potential enables the successful creation of hydrogenated liquid and amorphous silicon structures with realistic atomic configurations.
• The model demonstrates strong transferability across different hydrogenation levels and structural phases of silicon.
• Simulations using the GAP potential achieve significant computational speedups compared to direct DFT calculations while maintaining high accuracy.
• The validation against experimental data confirms the physical realism of the simulated hydrogenated silicon structures.