Systematic atomic structure datasets for machine learning potentials: Application to defects in magnesium

Aug 5, 2026

Upload by Logic

A physically motivated strategy constructs training sets for transferable machine learning interatomic potentials by systematically sampling all space groups in random crystal structures and applying controlled deformations to cell shape, size, and atomic positions. The resulting potentials remain unbiased and generically transferable to bulk defect studies without including defect configurations in training and without active learning. For pure magnesium, the method yields accurate potentials reproducing hcp and bcc polymorph properties, while analyses clarify how different training structure choices affect prediction quality and power.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Introduction

Defect engineering and the role of ML potentials

Data generation strategies: random sampling and active learning

Challenges in defect training data

Methodology and training set construction

Results for magnesium polymorphs and defect transferability

Impact of training structure choices

Systematic atomic structure datasets for machine learning potentials: Application to defects in magnesium Marvin Poul ,* Liam Huber,† Erik Bitzek ,‡ and Jörg Neugebauer§ Department of Computational Materials Design, Max-Planck-Institut für Eisenforschung GmbH, D-40327 Düsseldorf, Germany (Received 26 July 2022; revised 15 December 2022; accepted 8 February 2023; published 13 March 2023) We present a physically motivated strategy for the construction of training sets for transferable machine learning interatomic potentials. It is based on a systematic exploration of all possible space groups in random crystal structures, together with deformations of cell shape, size, and atomic positions. The resulting potentials turn out to be unbiased and generically applicable to studies of bulk defects without including any defect structures in the training set or employing any additional active learning. Using this approach we construct transferable potentials for pure magnesium that reproduce the properties of hexagonal closed packed (hcp) and body centered cubic (bcc) polymorphs very well. In the process we investigate how different types of training structures impact the properties and the predictive power of the resulting potential. DOI: 10.1103/PhysRevB.107.104103 I. INTRODUCTION A key concept in materials science to design materials with tailored properties is defect engineering. In order to successfully employ this concept, one needs a detailed understanding of the relationship between crystal defects on the atomistic scale and their inﬂuence on macroscopic materials properties. Until now this understanding has been provided to a large extent by density functional theory (DFT) calculations especially when investigating, e. g., the thermodynamic stability of materials phases and simple, isolated defects such as vacancies [1], dislocation arrays [2], or high-symmetry planar defects [3,4] . However, successful defect engineering must include most of the macroscopic and microscopic degrees of freedom of the defects—or risk missing potential candidate states. Especially in extended defects such as grain boundaries this defect phase space is very large, making it unfeasible to scan with DFT due to its high computational cost and system size restrictions. Together with recent interest in defect phase diagrams [5,6] this motivates us to develop a machine learning potential speciﬁcally aimed at a transferable description of defects. To this end, we will apply the moment tensor potential (MTP) methodology [7], and rigorously examine the impact of training data on the quality and performance of the resulting potentials. The approach and the detailed analysis and discussion are however general and can be applied to any machine learning (ML) potential methodology. *[poul@mpie.de](poul@mpie.de)[ ](poul@mpie.de)†[huber@mpie.de](huber@mpie.de)[ ](huber@mpie.de)‡[bitzek@mpie.de](bitzek@mpie.de) §[neugebauer@mpie.de](neugebauer@mpie.de) Published by the American Physical Society under the terms of the Creative Commons Attribution 4 .0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI. Open access publication funded by the Max Planck Society. Classical potentials are often trained on a set of properties that they ought to reproduce, e.g., relative phase stabilities, surface energies, and elastic properties. The more data hungry machine learning potentials instead use large sets of reference structure with energies, forces, and potentially stresses calculated with quantum mechanical models like DFT. These reference structures are generally constructed starting from equilibrium structures of interest, which are then perturbed in various ways to sample the energy landscape. This approach can work very well, but can lead to failure of the potential when relevant structures are missing. Another approach recently presented

FAQs

What is the key idea behind the training set construction method?

It uses systematic exploration of all space groups in random crystal structures combined with deformations of cell shape, size, and atomic positions to generate training data.

Does the method require defect structures in the training set or active learning?

No. The paper emphasizes that the potentials are unbiased and generically applicable to bulk defects without including any defect structures in training and without using additional active learning.

How do the resulting machine learning potentials perform for magnesium?

For pure magnesium, the constructed transferable potentials reproduce the properties of hexagonal close packed (hcp) and body centered cubic (bcc) polymorphs very well.

Aug 5, 2026

Upload by Logic