Scholarlink

<p>A deep-learning-based multi-target regression model was adopted to predict the mechanical properties of HRB500E high-strength seismic-resistant steel bars, incorporating key physical-metallurgical features. Three deep learning architectures were constructed and compared: a multitask learning network, a residual network, and an attention network. The results show that the attention network achieved the best overall performance across all evaluation metrics. On the training and test sets, it attained comprehensive coefficients of determination (<i>R</i><sup>2</sup>) of 0.9919 and 0.9711, comprehensive root mean square errors (RMSE) of 0.8420 and 1.6077, and comprehensive mean absolute errors (MAE) of 0.5104 and 0.7777, respectively, demonstrating excellent predictive stability. To interpret the attention network, SHapley Additive exPlanations (SHAP) analysis was conducted, confirming the consistent and significant contribution of physical-metallurgical features across all target predictions. By integrating physical metallurgy principles with deep learning, this study enables high-accuracy prediction of the mechanical properties of HRB500E hot-rolled ribbed steel bars, offering an innovative solution for intelligent steel manufacturing. This approach facilitates a paradigm shift in materials research from traditional empirical methods toward a data-mechanism fusion framework.</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

Multi-objective Prediction and Interpretability Analysis of Mechanical Properties for HRB500E Based on Deep Learning

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Ruifei Xuan,
  • Dazheng Zhang,
  • Tingfeng Xu,
  • Xingyu Wang,
  • Ziyi Zhao,
  • Weijuan Li

摘要

A deep-learning-based multi-target regression model was adopted to predict the mechanical properties of HRB500E high-strength seismic-resistant steel bars, incorporating key physical-metallurgical features. Three deep learning architectures were constructed and compared: a multitask learning network, a residual network, and an attention network. The results show that the attention network achieved the best overall performance across all evaluation metrics. On the training and test sets, it attained comprehensive coefficients of determination (R2) of 0.9919 and 0.9711, comprehensive root mean square errors (RMSE) of 0.8420 and 1.6077, and comprehensive mean absolute errors (MAE) of 0.5104 and 0.7777, respectively, demonstrating excellent predictive stability. To interpret the attention network, SHapley Additive exPlanations (SHAP) analysis was conducted, confirming the consistent and significant contribution of physical-metallurgical features across all target predictions. By integrating physical metallurgy principles with deep learning, this study enables high-accuracy prediction of the mechanical properties of HRB500E hot-rolled ribbed steel bars, offering an innovative solution for intelligent steel manufacturing. This approach facilitates a paradigm shift in materials research from traditional empirical methods toward a data-mechanism fusion framework.