Scholarlink

<p>The prediction of crystal lattice types from chemical composition remains a fundamental challenge in materials science and crystallography. Traditional experimental methods such as X-ray diffraction, while accurate, are inherently sequential processes that limit their applicability in high-throughput virtual screening of hypothetical compounds. This study presents <b>MGCSP-DL</b> (Materials Graph Crystal Lattice Prediction via Deep Learning), an approach that leverages Graph Neural Networks (GNNs) combined with advanced deep learning architectures to predict crystal lattice types across all seven crystal systems and 14 Bravais lattices directly from chemical composition. Using a comprehensive dataset of more than 389K materials from the Materials Project database, with thermodynamic stability filtering to retain only ground-state structures, our approach achieves classification accuracy ranging from 90.15% to 98.38% across different crystal systems. The approach demonstrates superior performance with 98.38% accuracy for cubic lattice types, 98.07% for hexagonal, 96.39% for trigonal, 95.64% for tetragonal, 91.99% for monoclinic, and 90.15% for orthorhombic systems. By encoding chemical compositions as weighted graphs with 33-dimensional element feature vectors and applying message-passing neural networks, MGCSP-DL provides an efficient computational pathway for rapid materials screening, achieving predictions in milliseconds compared to hours or days required by density functional theory calculations.</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

MGCSP-DL: a deep learning approach for material graph crystal lattice prediction

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Outhman Abbassi,
  • Hicham Labrim,
  • Soumia Ziti,
  • Mohammed Benaissa

摘要

The prediction of crystal lattice types from chemical composition remains a fundamental challenge in materials science and crystallography. Traditional experimental methods such as X-ray diffraction, while accurate, are inherently sequential processes that limit their applicability in high-throughput virtual screening of hypothetical compounds. This study presents MGCSP-DL (Materials Graph Crystal Lattice Prediction via Deep Learning), an approach that leverages Graph Neural Networks (GNNs) combined with advanced deep learning architectures to predict crystal lattice types across all seven crystal systems and 14 Bravais lattices directly from chemical composition. Using a comprehensive dataset of more than 389K materials from the Materials Project database, with thermodynamic stability filtering to retain only ground-state structures, our approach achieves classification accuracy ranging from 90.15% to 98.38% across different crystal systems. The approach demonstrates superior performance with 98.38% accuracy for cubic lattice types, 98.07% for hexagonal, 96.39% for trigonal, 95.64% for tetragonal, 91.99% for monoclinic, and 90.15% for orthorhombic systems. By encoding chemical compositions as weighted graphs with 33-dimensional element feature vectors and applying message-passing neural networks, MGCSP-DL provides an efficient computational pathway for rapid materials screening, achieving predictions in milliseconds compared to hours or days required by density functional theory calculations.