Local-environment-guided selection of atomic structures for the development of machine-learning potentials

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning potentials (MLPs) are vital in computational chemistry and materials science for their accuracy and efficiency.
• Developing reliable MLPs hinges on selecting appropriate atomic structures for training.
• Redundant or insufficient data hinders MLP development and performance.

Purpose of the Study:

• To propose a local-environment-guided screening algorithm for efficient dataset selection in MLP development.
• To reduce the size of training datasets while maintaining MLP accuracy.
• To enhance the robustness and computational efficiency of MLP model training.

Main Methods:

• A local environment bank stores unique atomic local environments.
• Dissimilarity is assessed using Euclidean distance to identify novel environments.
• New structures are selected only if their local environments differ significantly from existing ones.
• The bank is updated with new local environments from selected structures.

Main Results:

• The algorithm reduced training data size by approximately 80% for Ge and Pd13H2 systems.
• MLP model performance was not compromised despite the reduced dataset size.
• The method demonstrated superior robustness and computational efficiency compared to farthest point sampling.
• Results were independent of initial structure selection and ordering.

Conclusions:

• The proposed algorithm enables efficient dataset selection for developing accurate and computationally efficient MLPs.
• The local environment bank can serve as a continuously updated database for future MLP development.
• This approach optimizes the training data selection process, leading to better MLP models.