Comparison of intermediate-range order in GeO$_2$ glass: molecular dynamics using machine-learning interatomic potential vs.\ reverse Monte Carlo fitting to experimental data

TL;DR

This study compares GeO$_2$ glass structures obtained via machine-learning molecular dynamics and reverse Monte Carlo fitting, revealing significant differences in intermediate-range order, especially in ring distributions, due to the underlying modeling approaches.

Contribution

It demonstrates how different modeling techniques produce distinct structural features in GeO$_2$ glass, highlighting the impact of the density functional approximation and fitting constraints.

Findings

Similar two-body correlations in both models.

Different three-dimensional structures, especially in ring distributions.

Machine-learning MD yields narrower ring size distributions.

Abstract

The short and intermediate-range order in GeO$_2$ glass are investigated by molecular dynamics using machine-learning interatomic potential trained on ab initio calculation data and compared with reverse Monte Carlo fitting of neutron diffraction data. To characterize the structural differences in each model, the total/partial structure factors, coordination number, ring size and shape distributions, and persistent homology analysis were performed. These results show that although the two approaches yield similar two-body correlations, they can lead to three-dimensional models with very different short and intermediate-range ordering. A clear difference was observed especially in the ring distributions; RMC models exhibit a broad distribution in the ring size distribution, while neural network potential molecular dynamics yield much narrower ring distributions. This confirms that the density functional approximation in the ab initio calculations determines the preferred network assembly more strictly than RMC with simple coordination constraints and neutron diffraction data with isotope substitution.