Energy-free machine learning force field for aluminum

Area of Science:

• Materials Science
• Computational Physics
• Machine Learning

Background:

• Accurate interatomic potentials are crucial for molecular dynamics simulations.
• Developing reliable potentials for materials like aluminum and uranium is computationally challenging.

Purpose of the Study:

• To implement and validate a machine learning potential for molecular dynamics simulations.
• To assess the accuracy of the machine learning potential against experimental and ab initio data.
• To calculate key material properties using the developed potential.

Main Methods:

• Utilized the machine learning technique by Li et al. (PRL 114, 2015).
• Employed a feature matrix based on internal vectors to describe atomic configurations.
• Applied linear regression as the learning technique within the LAMMPS code.
• Tested the potential on crystalline and liquid aluminum and uranium.

Main Results:

• The machine learning potential demonstrated superior accuracy compared to existing potentials.
• Calculated phonon density of states, entropy, and melting temperature for aluminum.
• Achieved excellent agreement between predicted and experimental/ab initio results for aluminum properties.

Conclusions:

• The developed machine learning potential is a highly accurate and efficient tool for materials simulations.
• This approach provides a reliable method for predicting material properties.
• The findings support the use of machine learning in developing advanced interatomic potentials.