k-Means Clustering in Fingerprint-Based Configuration Selection for Fitting Interatomic Potentials.

Area of Science:

• Computational Materials Science
• Atomistic Simulations
• Machine Learning in Chemistry

Background:

• Accurate interatomic potentials are crucial for molecular dynamics simulations.
• Fitting potentials to quantum mechanical data (like DFT) requires a representative set of atomic configurations.
• Traditional methods often rely on random sampling, which can be inefficient.

Purpose of the Study:

• To develop a more efficient method for selecting representative atomic configurations for fitting interatomic potentials.
• To improve the accuracy and reduce the number of configurations needed for fitting potentials.
• To compare the efficacy of k-means clustering against random selection.

Main Methods:

• Applying k-means clustering to atomistic configuration fingerprints.
• Utilizing CrystalNN model and radial distribution function (RDF) for fingerprinting.
• Fitting embedded-atom method (EAM) potentials for titanium using selected configurations.
• Employing t-distributed stochastic neighbor embedding (t-SNE) for dimensionality reduction and visualization.

Main Results:

• K-means clustering significantly improves the accuracy of fitting interatomic potentials (energies and forces) compared to random selection.
• Fewer configurations (around 30) selected by k-means are sufficient to describe a larger dataset (1800 configurations).
• K-means clustering effectively identifies and utilizes configurations with similar atomic environments, even those with vacancies, which random selection misses.

Conclusions:

• K-means clustering offers a superior strategy for selecting configurations in materials modeling.
• This approach leads to more precise and reliable interatomic potentials with reduced computational cost.
• The findings suggest potential information redundancy in atomistic configurations, which k-means can exploit.