Evaluation of Machine Learning Interatomic Potentials for the Properties of Gold Nanoparticles

Area of Science:

• Computational Materials Science
• Nanotechnology
• Machine Learning

Background:

• Accurate simulation of nanomaterials like gold nanoparticles is crucial for understanding their properties.
• Traditional ab initio methods are computationally expensive, limiting large-scale simulations.
• Machine Learning Interatomic Potentials (ML-IPs) offer a promising alternative for efficient materials modeling.

Purpose of the Study:

• To investigate the application of ML-IPs, specifically using the DeePMD package, for simulating gold nanoparticle properties.
• To benchmark the accuracy and computational cost of ML-IPs against ab initio methods.
• To assess the capability of ML-IPs in reproducing structural and thermal properties of gold nanoclusters.

Main Methods:

• Utilized the DeePMD package for developing ML-IPs.
• Generated training data using the ab initio VASP program.
• Performed benchmarking simulations on Au20 nanoclusters using LAMMPS with ML-IPs.
• Compared ML-IP results with ab initio molecular dynamics simulations.

Main Results:

• Achieved accuracy comparable to ab initio molecular dynamics simulations using ML-IPs.
• Demonstrated significant reduction in computational cost with ML-IPs.
• Successfully reproduced structures and heat capacities of various isomeric forms of Au20 nanoclusters.

Conclusions:

• ML-IPs, particularly with DeePMD, provide a computationally efficient and accurate method for studying gold nanoparticle properties.
• The workflow shows potential for broader application in nanomaterials research.
• Identified areas for future improvement in ML-IP development and application.