Scholarlink

<p>This study utilizes three machine learning (ML) models—XGBoost, graph neural networks (GNN), and graph attention networks (GATs)—to predict unstable stacking fault energies (USFEs) in over 1000 refractory non-dilute random alloys, including mono, binary, ternary, quaternary, and quinary systems. Data for training are generated through atomistic simulations. In ML models, each chemical composition is encoded as a five-dimensional numerical vector to facilitate the representation of features, which are then normalized to balance the input scales during model training. Although explicit inter-composition connections are not included, the models learn compositional trends through node features and internal attention mechanisms. Because the lattice parameter is a byproduct in calculating the USFE in atomistic simulations, we also build three models for the lattice parameters. All models achieve reasonable accuracy in predicting lattice parameters but show variability in USFE predictions. Among the three models, GAT achieves the highest accuracy with average errors of 0.68% for USFE and 0.09% for the lattice parameter, outperforming GNN (1.24%, 0.11%) and XGBoost (0.46%, 0.54%). This study highlights the potential of graph-based ML models in predicting material properties from simulation-derived datasets, offering insights that could enhance traditional methods in materials science.</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

Integrating Atomistic Simulations and Machine Learning to Predict Unstable Stacking Fault Energies of Refractory Non-dilute Random Alloys

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Richard Brinlee,
  • Amin Poozesh,
  • Anvesh Nathani,
  • Anshu Raj,
  • Xiang-Guo Li,
  • Iman Ghamarian,
  • Shuozhi Xu

摘要

This study utilizes three machine learning (ML) models—XGBoost, graph neural networks (GNN), and graph attention networks (GATs)—to predict unstable stacking fault energies (USFEs) in over 1000 refractory non-dilute random alloys, including mono, binary, ternary, quaternary, and quinary systems. Data for training are generated through atomistic simulations. In ML models, each chemical composition is encoded as a five-dimensional numerical vector to facilitate the representation of features, which are then normalized to balance the input scales during model training. Although explicit inter-composition connections are not included, the models learn compositional trends through node features and internal attention mechanisms. Because the lattice parameter is a byproduct in calculating the USFE in atomistic simulations, we also build three models for the lattice parameters. All models achieve reasonable accuracy in predicting lattice parameters but show variability in USFE predictions. Among the three models, GAT achieves the highest accuracy with average errors of 0.68% for USFE and 0.09% for the lattice parameter, outperforming GNN (1.24%, 0.11%) and XGBoost (0.46%, 0.54%). This study highlights the potential of graph-based ML models in predicting material properties from simulation-derived datasets, offering insights that could enhance traditional methods in materials science.