Automatic selection of atomic fingerprints and reference configurations for machine-learning potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning (ML) potentials offer first-principles accuracy at reduced computational cost for molecular modeling.
• The performance of ML potentials heavily relies on the chosen atomic descriptors, which represent atomic configurations.
• Descriptors encode structural information and invariances (rotation, translation, permutation) crucial for potential energy surfaces.

Purpose of the Study:

• To develop automatic protocols for selecting optimal atomic descriptors from a large set of candidates.
• To simplify the construction of accurate and computationally efficient neural network potentials (NNPs).
• To accelerate the evaluation of Gaussian approximation potentials (GAP).

Main Methods:

• Automatic selection of atomic descriptors based on correlations within training data.
• Construction of neural network potentials (NNPs).
• Application of Gaussian process regression (GPR) for predicting formation energies.

Main Results:

• Demonstrated a procedure to simplify NNP construction by selecting relevant fingerprints.
• Achieved a balance between accuracy and computational efficiency in ML potentials.
• Showcased applications for water, Al-Mg-Si alloys, and organic molecule formation energies.

Conclusions:

• Automatic descriptor selection streamlines the development of high-performance ML potentials.
• The approach enhances both accuracy and computational efficiency in molecular modeling.
• This method accelerates the prediction of material and molecular properties.