Building machine learning force fields for nanoclusters.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics

Background:

• Modeling interatomic forces is crucial for understanding material properties.
• Gaussian Process (GP) regression offers a data-driven approach to force field development.
• Existing methods struggle with the complexity of nanocluster structures.

Purpose of the Study:

• To evaluate the efficacy of GP regression with varying kernel complexities (2-body, 3-body, many-body) for modeling interatomic forces in metal nanoclusters.
• To develop an efficient, non-parametric force field for predicting structural properties of nanoclusters at finite temperatures.
• To assess the thermal stability of 19-atom Nickel (Ni) nanoclusters.

Main Methods:

• Applied GP regression with 2-body, 3-body, and many-body kernels to Ni cluster structures.
• Trained and tested kernels on various datasets, including out-of-sample and high-temperature structures.
• Developed a non-parametric 3-body force field based on heterogeneous training data.
• Utilized the developed force field to predict thermal stability of Ni19 nanoclusters.

Main Results:

• Two-body GP kernels showed insufficient accuracy for nanoclusters, unlike in bulk systems.
• Three- and many-body kernels achieved high accuracy (∼0.1 eV/Å average error) even with small datasets.
• Extrapolation to dissimilar structures was challenging but improved with heterogeneous training.
• The developed 3-body force field accurately predicted structural properties at finite temperatures.

Conclusions:

• Advanced GP kernels (3- and many-body) are effective for modeling interatomic forces in metal nanoclusters.
• Heterogeneous training is key to overcoming extrapolation issues and improving model versatility.
• The developed non-parametric force field offers a computationally efficient alternative to ab initio methods for nanocluster simulations.