Search

We analyse the efficacy of machine learning (ML) interatomic potentials (IP) in modelling gold (Au) nanoparticles. We have explored the transferability of these ML models to larger systems and established simulation times and size thresholds necessary for accurate interatomic potentials. To achieve this, we compared the energies and geometries of large Au nanoclusters using VASP and LAMMPS and gained better understanding of the number of VASP simulation timesteps required to generate ML-IPs that can reproduce the structural properties. We also investigated the minimum atomic size of the training set necessary to construct ML-IPs that accurately replicate the structural properties of large Au nanoclusters, using the LAMMPS-specific heat of the Au147 icosahedral as reference. Our findings suggest that minor adjustments to a potential developed for one system can render it suitable for other systems. These results provide further insight into the development of accurate interatomic potentials for modelling Au nanoparticles through machine learning techniques.

We have investigated Machine Learning Interatomic Potentials in application to the properties of gold nanoparticles through the DeePMD package, using data generated with the ab-initio VASP program. Benchmarking was carried out on Au20 nanoclusters against ab-initio molecular dynamics simulations and show we can achieve similar accuracy with the machine learned potential at far reduced cost using LAMMPS. We have been able to reproduce structures and heat capacities of several isomeric forms. Comparison of our workflow with similar ML-IP studies is discussed and has identified areas for future improvement.