Overcoming the Chemical Complexity Bottleneck in on-the-Fly Machine Learned Molecular Dynamics Simulations

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Molecular dynamics (MD) simulations are crucial for understanding material properties.
• Accurate force fields are essential for reliable MD simulations.
• Current methods face computational bottlenecks with increasing chemical complexity.

Purpose of the Study:

• To develop an on-the-fly machine learning force field framework for MD simulations.
• To overcome the computational limitations associated with a large number of chemical elements.
• To enable efficient simulations of complex bulk systems.

Main Methods:

• Utilized a multipole featurization scheme for machine learning force fields.
• Implemented an on-the-fly learning approach within MD simulations.
• Tested the framework on bulk systems containing up to six chemical elements.

Main Results:

• The number of density functional theory (DFT) calls remained largely independent of the number of chemical elements.
• This approach significantly overcomes the bottleneck seen in other methods like the smooth overlap of atomic positions (SOAP) scheme.
• Demonstrated scalability for systems with diverse elemental compositions.

Conclusions:

• The multipole featurization framework offers a computationally efficient solution for ML-based MD simulations.
• This method enables the study of complex materials with multiple elements.
• Provides a pathway for accelerating materials discovery and design.