Starts 8 Oct 2025 11:00
Ends 8 Oct 2025 12:00
Central European Time
CMSP Seminar (Atomistic Simulation Seminar Series): Inference of classical partial charges distribution for molecular dynamics
Vojtech Kostal
(Czech Academy of Sciences,
Institute of Organic Chemistry and Biochemistry)
(Czech Academy of Sciences,
Institute of Organic Chemistry and Biochemistry)
Abstract:
Molecular dynamics simulations rely on empirical fixed-charge force fields to achieve atomistic resolution at biologically relevant scales. Traditional parameterization pipelines typically yield a single “best” parameter set, offering little flexibility for refinement and leaving users without guidance on safe ranges for modifications. Here, we present a Bayesian framework for optimizing partial charges anchored to condensed-phase reference data from density functional theory molecular dynamics. Structural quantities of interest, such as radial distribution functions and hydrogen bonding patterns, are used as optimization targets, while a local Gaussian process surrogate accelerates sampling in high-dimensional charge space. This approach yields posterior distributions of parameters, quantifying uncertainty and enabling systematic refinement beyond a single solution. Applied to a diverse set of neutral and charged molecules, our method establishes a transferable, robust, and physically grounded workflow for improving fixed-charge force fields. To account for the inherently missing electronic polarization, we employ the electronic continuum correction as a standard underlying the parameterization of charged species.
Molecular dynamics simulations rely on empirical fixed-charge force fields to achieve atomistic resolution at biologically relevant scales. Traditional parameterization pipelines typically yield a single “best” parameter set, offering little flexibility for refinement and leaving users without guidance on safe ranges for modifications. Here, we present a Bayesian framework for optimizing partial charges anchored to condensed-phase reference data from density functional theory molecular dynamics. Structural quantities of interest, such as radial distribution functions and hydrogen bonding patterns, are used as optimization targets, while a local Gaussian process surrogate accelerates sampling in high-dimensional charge space. This approach yields posterior distributions of parameters, quantifying uncertainty and enabling systematic refinement beyond a single solution. Applied to a diverse set of neutral and charged molecules, our method establishes a transferable, robust, and physically grounded workflow for improving fixed-charge force fields. To account for the inherently missing electronic polarization, we employ the electronic continuum correction as a standard underlying the parameterization of charged species.