Spin parameter optimization for spin-polarized extended tight-binding methods

TL;DR

This paper introduces an optimization strategy for atom-specific spin-polarization constants in the GFN2-xTB method, significantly improving accuracy for certain datasets but revealing limited transferability across different molecular properties.

Contribution

The study develops and compares sequential and global optimization methods for spin parameters, using sensitivity analysis to enhance molecular simulation accuracy within the extended tight-binding framework.

Findings

Error reduction on the W4-11 dataset

Limited transferability of optimized parameters across properties

Inherent limitations of current methods not fixed by parameter optimization

Abstract

We present an optimization strategy for atom-specific spin-polarization constants within the spin-polarized GFN2-xTB framework, aiming to enhance the accuracy of molecular simulations. We compare a sequential and global optimization of spin parameters for hydrogen, carbon, nitrogen, oxygen, and fluorine. Sensitivity analysis using Sobol indices guides the identification of the most influential parameters for a given reference dataset, allowing for a nuanced understanding of their impact on diverse molecular properties. In the case of the W4-11 dataset, substantial error reduction was achieved, demonstrating the potential of the optimization. Transferability of the optimized spin-polarization constants over different properties, however, is limited, as we demonstrate by applying the optimized parameters on a set of singlet-triplet gaps in carbenes. Further studies on ionization potentials and electron affinities highlight some inherent limitations of current extended tight-binding methods that can not be resolved by simple parameter optimization. We conclude that the significantly improved accuracy strongly encourages the present re-optimization of the spin-polarization constants, whereas the limited transferability motivates a property-specific optimization strategy.