A neural network-assisted open boundary molecular dynamics simulation method.

Area of Science:

• Computational chemistry
• Materials science
• Statistical mechanics

Background:

• Molecular dynamics (MD) simulations are crucial for understanding material properties.
• Simulating open boundary systems, which mimic real-world conditions, is computationally expensive.
• Existing methods struggle to efficiently model the effects of unmodeled surrounding fluids.

Purpose of the Study:

• To develop a computationally efficient method for open boundary molecular dynamics simulations.
• To accurately represent the influence of surrounding fluids on a simulated system.
• To reduce the significant computational cost associated with traditional open boundary simulations.

Main Methods:

• A neural network-assisted molecular dynamics approach was developed.
• Particle influxes and neural network-derived forces were applied at the simulation domain boundaries.
• Canonical ensemble simulations with periodic boundaries were used for neural network training and boundary flux sampling.
• The method was implemented within the LAMMPS simulation package.

Main Results:

• The method achieved results within 2.5% accuracy for temperature, kinetic energy, potential energy, and pressure compared to periodic simulations.
• The new approach demonstrated a speedup of two orders of magnitude over comparable grand canonical molecular dynamics simulations.
• Accurate representation of surrounding fluid effects was achieved through boundary modeling.

Conclusions:

• The neural network-assisted method significantly reduces computational cost for open boundary simulations.
• This technique offers a viable and efficient alternative for simulating systems with external fluid interactions.
• The approach shows promise for broader applications in computational chemistry and materials science.