Bayesian-Inference-Driven Model Parametrization and Model Selection for 2CLJQ Fluid Models.

Area of Science:

• Computational chemistry
• Statistical mechanics
• Molecular modeling

Background:

• High physical detail in molecular models enhances simulation accuracy but increases computational cost.
• Model complexity must be balanced against computational feasibility and the need to capture specific properties.

Purpose of the Study:

• To apply Bayesian inference and surrogate modeling for selecting optimal complexity in molecular models.
• To evaluate the Bayes factor evidence for different complexity levels of the two-centered Lennard-Jones + quadrupole (2CLJQ) fluid model.
• To analyze the impact of parameter priors on model selection.

Main Methods:

• Utilized Bayesian inference for molecular model selection.
• Employed Monte Carlo sampling techniques accelerated with surrogate modeling.
• Evaluated Bayes factor evidence for three nested levels of complexity in the 2CLJQ model.

Main Results:

• Determined that variable quadrupole and bond length parameters are justified only for certain chemistries.
• Obtained detailed information on parameter distributions and correlations, aiding parametrization.
• Demonstrated the significant effect of parameter priors on model selection outcomes.

Conclusions:

• Bayesian inference provides a robust framework for molecular model selection, balancing accuracy and computational cost.
• The study offers a computational roadmap for applying these techniques to future molecular modeling challenges.
• Careful consideration of parameter priors is crucial to avoid unnecessary model complexity.