Bayesian Modeling of Polarizable Water: Lessons for Force Field Development.

Area of Science:

• Computational chemistry and materials science.
• Development of accurate molecular models for simulations.

Background:

• Molecular dynamics simulations require accurate models and quantified prediction uncertainties.
• Current methods often neglect parameter uncertainties, impacting model reliability.
• Previous work developed Bayesian three-point water models, highlighting dependence on reference observables.

Purpose of the Study:

• Extend Bayesian inference to polarizable water models (SWM4-NDP).
• Investigate the impact of different van der Waals functional forms and inference observables.
• Assess the role of explicit polarization in reducing parameter uncertainty.

Main Methods:

• Applied Bayesian inference and Markov-chain Monte Carlo sampling to the SWM4-NDP polarizable water model.
• Compared two van der Waals functional forms (Lennard-Jones 12-6 and Wang-Buckingham).
• Trained models using gas-phase dimer energies and compared with Bayesian inference results.

Main Results:

• Two van der Waals functional forms showed similar performance but favored different observables.
• Explicit polarization did not significantly reduce parameter uncertainty for the selected observables.
• Bayesian inference yielded optimal parameters similar to those from gas-phase dimer training for Lennard-Jones 12-6.
• Wang-Buckingham potential trained on gas-phase data was unstable, unlike the Bayesian counterpart.

Conclusions:

• Force field parameter uncertainty quantification is essential for reliable molecular simulations.
• A diverse set of observables is necessary for robust force field development.
• Bayesian inference provides a robust framework for developing accurate and reliable polarizable water models.