Machine learning-enhanced multiple time-step ab initio molecular dynamics.

Area of Science:

• Computational Chemistry
• Materials Science
• Physical Chemistry

Background:

• Molecular dynamics (MD) simulations are computationally expensive.
• The efficiency of MD is limited by the time step, dictated by high-frequency motions.
• Multiple time step (MTS) methods address this by using different time steps for fast and slow force components.

Purpose of the Study:

• To develop a machine learning-enhanced multiple time step (ML-MTS) method.
• To achieve accurate Born-Oppenheimer molecular dynamics at reduced computational cost.
• To present two novel ML-MTS schemes for stable and accurate simulations.

Main Methods:

• Developed two ML-MTS schemes integrating machine learning force fields.
• Scheme 1: ML force estimates replace high-level calculations.
• Scheme 2: ML correction applied to fast components with high-level calculation for slow components.

Main Results:

• Scheme 1 achieved speedups of two orders of magnitude over standard velocity Verlet (VV) integration.
• Scheme 2 enabled a fourfold increase in time step compared to ab initio MTS algorithms.
• Both schemes yielded stable and accurate trajectories, with Scheme 2 providing speedups up to an order of magnitude over VV.

Conclusions:

• ML-MTS methods offer a significant reduction in computational cost for accurate molecular dynamics.
• The proposed schemes provide stable and efficient alternatives to traditional MTS and VV integration.
• This work paves the way for more accessible and faster large-scale molecular dynamics simulations.