Improving Molecular Dynamics of Aqueous Electrolytes through Data-Driven Sensitivity-Guided GFN2-xTB Reparameterization
8 affiliations
- 1Department of Chemistry, TUM School of Natural Sciences
- 2Technical University of Munich
- 3Department of Chemical Engineering
- 4Massachusetts Institute of Technology
- 5Institute for Advanced Study, Technical University of Munich, Lichtenbergstrasse 2 a, Garching, Germany, 85748
- 6Department of Chemistry
- 7Catalysis Research Center
- 8Munich Data Science Institute
Abstract
Semiempirical tight-binding methods such as GFN2-xTB enable efficient molecular dynamics (MD) simulations, but their performance is strongly system dependent. The default GFN2-xTB parameterization overestimates equilibrium densities of liquid water and aqueous NaCl, and it underestimates the bond length for Na+–water pairs in the radial distribution function. To correct these deficiencies, we develop a sensitivity-guided stepwise approach to reparameterize GFN2-xTB against cluster-based and periodic revPBE0-D3 data. We find that the GFN2-xTB fitting problem is ill-conditioned, with many compensating parameter changes that reduce fitted errors but yield unphysical densities in MD simulations. Fisher information matrix (FIM) analysis reveals that only a few parameter combinations dominate the sensitivity of fitted observables, suggesting a reduction of the parameter space for efficient optimization. Optimizing those FIM-selected parameter subsets first for H/O and then for Na improves the equilibrium densities and shifts Na+ hydration distances into the experimental ranges without compromising the remaining properties. We achieve these improvements without directly fitting densities or hydration distances. While the best agreement with experiment for densities is achieved by optimizations on periodic datasets that include virials in the objective function, inclusion of targeted scans of essential geometric degrees of freedom in either a periodic or cluster dataset also improves optimization. Our results identify cluster data as a practical route to correct ion-solvation structure via inexpensive reference calculations, while periodic virials provide complementary control over bulk properties. These findings demonstrate that sensitivity-guided parameter selection, combined with carefully selected reference data describing the relevant interactions, enables GFN2-xTB reparameterizations whose improvements in the static fit translate to bulk MD and provides a general strategy for improving dynamical properties from low-cost electronic structure methods.
Related
- Machine-Learning Molecular Dynamics Reveals Grotthuss Transport in Concentrated Electrolytes
- When Is Molecular-Dynamics-Predicted Ionic Conductivity Reliable in Solid Electrolytes?
- Two-Dimensional IR Spectroscopy as a Window into the Molecular Structure and Dynamics of Electrolytes and Ionic Liquids: A Review
- Low-Temperature Transport in Li-Ion Battery EC/EMC/FEC Electrolytes: Molecular Dynamics and Machine-Learning Modeling
- Length-scale dependence of collective time reparameterization in swap dynamics
BibTeX
@misc{nikolaeva2026improving,
title = {{Improving Molecular Dynamics of Aqueous Electrolytes through Data-Driven Sensitivity-Guided GFN2-xTB Reparameterization}},
author = {Tatiana Nikolaeva and Heather J. Kulik and Christopher J. Stein},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010774/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010774/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Improving Molecular Dynamics of Aqueous Electrolytes through Data-Driven Sensitivity-Guided GFN2-xTB Reparameterization
AU - Tatiana Nikolaeva
AU - Heather J. Kulik
AU - Christopher J. Stein
PY - 2026
DA - 2026/10/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010774/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010774/v1
AB - Semiempirical tight-binding methods such as GFN2-xTB enable efficient molecular dynamics (MD) simulations, but their performance is strongly system dependent. The default GFN2-xTB parameterization overestimates equilibrium densities of liquid water and aqueous NaCl, and it underestimates the bond length for Na+–water pairs in the radial distribution function. To correct these deficiencies, we develop a sensitivity-guided stepwise approach to reparameterize GFN2-xTB against cluster-based and periodic revPBE0-D3 data. We find that the GFN2-xTB fitting problem is ill-conditioned, with many compensating parameter changes that reduce fitted errors but yield unphysical densities in MD simulations. Fisher information matrix (FIM) analysis reveals that only a few parameter combinations dominate the sensitivity of fitted observables, suggesting a reduction of the parameter space for efficient optimization. Optimizing those FIM-selected parameter subsets first for H/O and then for Na improves the equilibrium densities and shifts Na+ hydration distances into the experimental ranges without compromising the remaining properties. We achieve these improvements without directly fitting densities or hydration distances. While the best agreement with experiment for densities is achieved by optimizations on periodic datasets that include virials in the objective function, inclusion of targeted scans of essential geometric degrees of freedom in either a periodic or cluster dataset also improves optimization. Our results identify cluster data as a practical route to correct ion-solvation structure via inexpensive reference calculations, while periodic virials provide complementary control over bulk properties. These findings demonstrate that sensitivity-guided parameter selection, combined with carefully selected reference data describing the relevant interactions, enables GFN2-xTB reparameterizations whose improvements in the static fit translate to bulk MD and provides a general strategy for improving dynamical properties from low-cost electronic structure methods.
ER -