Scholarlink

<p>Self-piercing riveting (SPR) is increasingly employed for joining lightweight hybrid structures, particularly when traditional fusion welding is unsuitable for materials such as fibre-reinforced composites. However, the rivet-and-die design in SPR remains challenging due to the complex deformation and failure mechanisms involved in composite-metal joining. This study develops a hybrid, data-driven SPR rivet-and-die design framework that integrates machine learning (ML) with numerical simulations. The framework is demonstrated using the SPR of Glass Fibre-Reinforced Polymer (GFRP) composite sheets and Drawing Quality (DQ) steel sheets, illustrating their potential as dissimilar materials for automotive applications. SPR experiments are performed to validate numerical results and establish a reliable numerical simulation approach. The validated simulation setup is then used to perform multiple simulations with variations in rivet length, diameter, hardness, die diameter, die depth, pip height, friction coefficient, and sheet thickness, generating a comprehensive dataset. Additionally, parameter variations are generated using the Conditional Tabular Generative Adversarial Network (CTGAN) to produce statistically balanced synthetic data for the simulation input. The framework leverages simulation data to train predictive ML models and is validated by using experimental data that determine optimal die dimensions for various GFRP/DQ steel thickness combinations. The developed ML framework achieved strong predictive capability, with FE–ML correlation coefficients of <i>r</i> = 0.90, 0.95, and 0.89 for ID, BRT, and CD, respectively. The optimized design predicted by the Random Forest–L-BFGS-B framework showed close agreement with FE validation results, with deviations of only 1.1% for ID and 5.8% for BRT. The ultimate goal is to reduce experimental effort and design iterations by enabling accurate die prediction through the combined use of simulation and data-driven modelling. This approach is expected to be applicable to other hybrid joining and forming processes, as well as to other dissimilar sheet combinations that require efficient, interpretable design development.</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

Machine learning-based rivet–die design framework for self-piercing riveting of GFRP composite and DQ steel sheets

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Durga Rao Boddepalli,
  • R. Ganesh Narayanan,
  • B. R. Manoj,
  • Brajesh Asati,
  • Ishwar Kapoor

摘要

Self-piercing riveting (SPR) is increasingly employed for joining lightweight hybrid structures, particularly when traditional fusion welding is unsuitable for materials such as fibre-reinforced composites. However, the rivet-and-die design in SPR remains challenging due to the complex deformation and failure mechanisms involved in composite-metal joining. This study develops a hybrid, data-driven SPR rivet-and-die design framework that integrates machine learning (ML) with numerical simulations. The framework is demonstrated using the SPR of Glass Fibre-Reinforced Polymer (GFRP) composite sheets and Drawing Quality (DQ) steel sheets, illustrating their potential as dissimilar materials for automotive applications. SPR experiments are performed to validate numerical results and establish a reliable numerical simulation approach. The validated simulation setup is then used to perform multiple simulations with variations in rivet length, diameter, hardness, die diameter, die depth, pip height, friction coefficient, and sheet thickness, generating a comprehensive dataset. Additionally, parameter variations are generated using the Conditional Tabular Generative Adversarial Network (CTGAN) to produce statistically balanced synthetic data for the simulation input. The framework leverages simulation data to train predictive ML models and is validated by using experimental data that determine optimal die dimensions for various GFRP/DQ steel thickness combinations. The developed ML framework achieved strong predictive capability, with FE–ML correlation coefficients of r = 0.90, 0.95, and 0.89 for ID, BRT, and CD, respectively. The optimized design predicted by the Random Forest–L-BFGS-B framework showed close agreement with FE validation results, with deviations of only 1.1% for ID and 5.8% for BRT. The ultimate goal is to reduce experimental effort and design iterations by enabling accurate die prediction through the combined use of simulation and data-driven modelling. This approach is expected to be applicable to other hybrid joining and forming processes, as well as to other dissimilar sheet combinations that require efficient, interpretable design development.