OpenAlex ·

Cluster-based Structural Similarity for Dataset Visualization and Data Selection for Machine Learning Interatomic Potentials

Yuto Iwasaki and Miguel A. Caro

Published
Source
OpenAlex
Published in
arXiv (Cornell University)
DOI
10.48550/arxiv.2609.39984
Original
Read it at the publisher
Topics
Materials ScienceStatistical MechanicsAtomic PhysicsOptics

Abstract

Machine learning interatomic potentials (MLIPs) are essential components for accelerating simulation-driven materials design. Data-efficient MLIP training relies on data-selection strategies that maximize structural diversity while limiting computationally expensive first-principles calculations. A key challenge in such strategies is evaluating structural similarity, which involves a trade-off between retaining information on individual atomic environments and reducing computational cost. Here, we propose similarity evaluation methods that achieve both representational fidelity and computational efficiency. Our method represents each structure using a small set of characteristic atomic environments identified by k-medoids clustering and computes pairwise similarity through optimal matching between these representatives or their distributions. Molecular benchmarks demonstrate that our method is approximately 50 times faster than the baseline method while also more clearly distinguishing structures with different chemical compositions. Similarity-based data-selection benchmarks demonstrate that our methods improve the data efficiency and stability of force prediction in MLIPs.

The authors’ abstract, as arXiv lists it. arXiv’s listings, abstracts included, are free for anyone to reuse.

Read it on ReadSearch

Each opens after you create an account, and brings you back to this paper.

Already a reader? Sign in

All new papers in Materials Science

Get the few papers worth your time in Materials Science every morning.

ReadSearch reads the new papers in your field each day and brings you the few worth reading, each with its key finding in a sentence.

Papers belong to their authors and publishers; every page here links to the original. All research topics