An On-the-Fly Approach to Construct Generalized Energy-Based Fragmentation Machine Learning Force Fields of Complex...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Developing accurate force fields for large, complex molecular systems is computationally demanding.
• Existing methods often require extensive data selection and parameter optimization.
• Efficient and automated approaches are needed for simulating diverse chemical systems.

Purpose of the Study:

• To develop an on-the-fly fragment-based machine learning (ML) approach for constructing ML force fields.
• To enable the accurate simulation of large and complex molecular systems.
• To automate the generation of force fields without manual data selection or parameter tuning.

Main Methods:

• Implemented an on-the-fly fragment-based ML approach combining subsystems.
• Utilized a nonparametric Gaussian process (GP) model for automatic force field generation.
• Employed the generalized energy-based fragmentation (GEBF) method for system decomposition.

Main Results:

• Constructed a GEBF-ML force field for a normal alkane (C60H122).
• Performed long-time molecular dynamics (MD) simulations showing favorable agreement with quantum mechanics (QM) for energy, forces, and dipole moments.
• Achieved excellent agreement in predicted IR spectra compared to ab initio MD results.

Conclusions:

• The GEBF-ML method provides an automatic and efficient approach to build ML force fields.
• This method is applicable to a broad range of complex systems, including biomolecules and supramolecular systems.
• Demonstrated the potential for accurate and rapid simulations of large molecular systems.