Scholarlink

<p>This study investigates the feasibility of using a pre-trained residual neural network (ResNet) deep convolutional neural network (DCNN) architecture to classify crystal structures of materials in the cubic system based on selected area electron diffraction patterns (SADPs), without relying on additional features. A labeling scheme grounded in geometric attributes, complying with 2D crystallography and thereby interpretable by both human crystallographers and computer vision, was applied to SADPs spanning all component systems, space groups, chemical formulae, and unit cell configurations within the cubic system domain. The resulting compilation of SADP label permutations suggested that nearly all cubic space groups can be identified from a set of zone-axis-specific SADPs, provided accurate label classification and permutation matching. Initial training with broad SADP image datasets, generated by transmission electron microscopy (TEM) simulation, revealed limited generalization despite reasonable test accuracy, attributed to wide scatter in the distribution of samples and variance in the geometric complexity present in SADPs across component systems, formula types, and unit cell prototypes. To address these bias and variance issues, dataset augmentation strategies were introduced, reflecting operational conditions of laboratory TEM. Additionally, an ensemble learning approach was implemented, training six independent ResNet models tailored to each component system. This strategy not only improved classification accuracy ranging from 96.50% to 99.59% but also enabled prediction of the number of constituent elements in the materials under investigation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Determination of space groups and component systems from selected area electron diffraction patterns by deep learning

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Jaemin Jeong,
  • Moonsoo Ra,
  • Jinha Jeong,
  • Seungsoon Jang,
  • Woong Lee

摘要

This study investigates the feasibility of using a pre-trained residual neural network (ResNet) deep convolutional neural network (DCNN) architecture to classify crystal structures of materials in the cubic system based on selected area electron diffraction patterns (SADPs), without relying on additional features. A labeling scheme grounded in geometric attributes, complying with 2D crystallography and thereby interpretable by both human crystallographers and computer vision, was applied to SADPs spanning all component systems, space groups, chemical formulae, and unit cell configurations within the cubic system domain. The resulting compilation of SADP label permutations suggested that nearly all cubic space groups can be identified from a set of zone-axis-specific SADPs, provided accurate label classification and permutation matching. Initial training with broad SADP image datasets, generated by transmission electron microscopy (TEM) simulation, revealed limited generalization despite reasonable test accuracy, attributed to wide scatter in the distribution of samples and variance in the geometric complexity present in SADPs across component systems, formula types, and unit cell prototypes. To address these bias and variance issues, dataset augmentation strategies were introduced, reflecting operational conditions of laboratory TEM. Additionally, an ensemble learning approach was implemented, training six independent ResNet models tailored to each component system. This strategy not only improved classification accuracy ranging from 96.50% to 99.59% but also enabled prediction of the number of constituent elements in the materials under investigation.