Modeling crystal defects using defect-informed neural networks

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yang, Ziduo, Liu, Xiaoqing, Zhang, Xiuying, Huang, Pengru, Novoselov, Kostya S., Shen, Lei
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908389168119808
author Yang, Ziduo
Liu, Xiaoqing
Zhang, Xiuying
Huang, Pengru
Novoselov, Kostya S.
Shen, Lei
author_facet Yang, Ziduo
Liu, Xiaoqing
Zhang, Xiuying
Huang, Pengru
Novoselov, Kostya S.
Shen, Lei
contents Most AI-for-Materials research to date has focused on ideal crystals, whereas real-world materials inevitably contain defects that play a critical role in modern functional technologies. The defects break geometric symmetry and increase interaction complexity, posing particular challenges for traditional ML models. Here, we introduce Defect-Informed Equivariant Graph Neural Network (DefiNet), a model specifically designed to accurately capture defect-related interactions and geometric configurations in point-defect structures. DefiNet achieves near-DFT-level structural predictions in milliseconds using a single GPU. To validate its accuracy, we perform DFT relaxations using DefiNet-predicted structures as initial configurations and measure the residual ionic steps. For most defect structures, regardless of defect complexity or system size, only 3 ionic steps are required to reach the DFT-level ground state. Finally, comparisons with scanning transmission electron microscopy (STEM) images confirm DefiNet's scalability and extrapolation beyond point defects, positioning it as a valuable tool for defect-focused materials research.
format Preprint
id arxiv_https___arxiv_org_abs_2503_15391
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Modeling crystal defects using defect-informed neural networks
Yang, Ziduo
Liu, Xiaoqing
Zhang, Xiuying
Huang, Pengru
Novoselov, Kostya S.
Shen, Lei
Materials Science
Computational Physics
Most AI-for-Materials research to date has focused on ideal crystals, whereas real-world materials inevitably contain defects that play a critical role in modern functional technologies. The defects break geometric symmetry and increase interaction complexity, posing particular challenges for traditional ML models. Here, we introduce Defect-Informed Equivariant Graph Neural Network (DefiNet), a model specifically designed to accurately capture defect-related interactions and geometric configurations in point-defect structures. DefiNet achieves near-DFT-level structural predictions in milliseconds using a single GPU. To validate its accuracy, we perform DFT relaxations using DefiNet-predicted structures as initial configurations and measure the residual ionic steps. For most defect structures, regardless of defect complexity or system size, only 3 ionic steps are required to reach the DFT-level ground state. Finally, comparisons with scanning transmission electron microscopy (STEM) images confirm DefiNet's scalability and extrapolation beyond point defects, positioning it as a valuable tool for defect-focused materials research.
title Modeling crystal defects using defect-informed neural networks
topic Materials Science
Computational Physics
url https://arxiv.org/abs/2503.15391