Improving the Silicon Interactions of GFN-xTB

TL;DR

This paper improves the GFN-xTB method's accuracy for silicon-containing systems by re-fitting silicon parameters using a large reference dataset, enhancing predictions of energies, forces, and geometries.

Contribution

The authors re-parametrize the GFN-xTB model specifically for silicon, significantly enhancing its accuracy for organosilicon compounds.

Findings

Improved accuracy in energy predictions for silicon compounds

Enhanced geometry and force predictions with the new parameters

Better performance in systems containing silicon

Abstract

A general-purpose Density Functional Tight Binding method, the GFN-xTB model is gaining increased popularity in accurate simulations that are out of scope for conventional ab initio formalisms. We show that in its original GFN1-xTB parametrization, organosilicon compounds are described poorly. This issue is addressed by re-fitting the model's silicon parameters to a data set of ten thousand reference compounds, geometry-optimized with the revPBE functional. The resulting GFN1-xTB-Si parametrization shows improved accuracy in the prediction of system energies, nuclear forces and geometries and should be considered for all applications of the GFN-xTB Hamiltonian to systems that contain silicon.