A machine learning potential construction based on radial distribution function sampling.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate reference data is vital for constructing reliable machine learning potentials (MLPs).
• Insufficient local configurations in training data can cause unphysical behavior in MLP-based molecular dynamics (MLP-MD) simulations.

Purpose of the Study:

• To develop an on-the-fly sampling method for reference data to enhance MLP construction.
• To improve the accuracy and robustness of MLPs for molecular dynamics simulations, particularly for water systems.

Main Methods:

• Proposed a novel radial distribution function (RDF)-based data sampling technique for on-the-fly reference data collection.
• Detected and extracted anomalous configurations from MLP-MD trajectories by analyzing RDF shapes.
• Integrated these structures into the reference dataset to refine the MLP.

Main Results:

• MLP-MD simulations utilizing the new sampling method produced trajectories with physically realistic features, including accurate RDF shapes and angle distributions.
• The refined MLPs demonstrated robustness, accurately simulating bulk water systems from molecular cluster data.
• Unphysical behaviors observed in simulations without the RDF-based sampling were effectively mitigated.

Conclusions:

• The RDF-based data sampling approach is a highly effective strategy for constructing accurate and robust MLPs.
• This method enables reliable extrapolation from small molecular systems to larger, bulk systems without requiring specialized expertise.
• The technique significantly improves the quality of MLP-MD simulations, aligning results with ab initio calculations.