Scholarlink

<p>Designing self-centering moment-resisting frames (SC-MRFs) requires extensive nonlinear time history analyses (NTHA), making the process computationally demanding. This study proposes an automated framework that combines finite element modeling, machine learning (ML), and active learning (AL) to efficiently develop surrogate models for SC-MRF seismic design. A Latin Hypercube Sampling method was employed to generate a diverse set of self-centering connection designs, which were iteratively refined using a two-step AL process to select the most informative samples for NTHA. These selected designs were analyzed under ground motions corresponding to three seismic hazard levels, yielding engineering demand parameters (EDPs) as response labels. Three groups of input variables were defined to construct data sets for predicting and interpreting story-level structural responses. These data sets were further merged to form a comprehensive multi-story SC-MRF data set. Two novel metaheuristic algorithms—Electric Eel Foraging and Puma Optimizer—along with three conventional optimization techniques, were integrated with ensemble learning approaches for hyperparameter tuning. Hybrid models were assessed using five performance metrics and validated against 18 newly generated SC-MRF designs. Feature importance analysis using Shapley Additive Explanations (SHAP) identified key parameters influencing EDPs, while multi-gene genetic programming (MGGP) produced explicit predictive equations to support performance-based design (PBD) of SC-MRFs. Results demonstrated that the AL-driven approach reduced the required number of training samples by 70% while maintaining at least 88% accuracy, outperforming models trained on 90 randomly selected samples. The best hybrid models achieved accuracy of 94%, 98.6%, and 99% in predicting story-level residual drift, acceleration, and maximum drift across all SC-MRF configurations. Furthermore, using MGGP-based equations, accurate performance was achieved with a minimum of 91% precision in predicting story-level EDP.</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

A Machine Learning Framework for Efficient Seismic Design of Self-Centering Steel Frames Using Active Learning and Surrogate Modeling

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Alireza Farzinpour,
  • Vahid Broujerdian,
  • Esmaeil Mohammadi Dehcheshmeh

摘要

Designing self-centering moment-resisting frames (SC-MRFs) requires extensive nonlinear time history analyses (NTHA), making the process computationally demanding. This study proposes an automated framework that combines finite element modeling, machine learning (ML), and active learning (AL) to efficiently develop surrogate models for SC-MRF seismic design. A Latin Hypercube Sampling method was employed to generate a diverse set of self-centering connection designs, which were iteratively refined using a two-step AL process to select the most informative samples for NTHA. These selected designs were analyzed under ground motions corresponding to three seismic hazard levels, yielding engineering demand parameters (EDPs) as response labels. Three groups of input variables were defined to construct data sets for predicting and interpreting story-level structural responses. These data sets were further merged to form a comprehensive multi-story SC-MRF data set. Two novel metaheuristic algorithms—Electric Eel Foraging and Puma Optimizer—along with three conventional optimization techniques, were integrated with ensemble learning approaches for hyperparameter tuning. Hybrid models were assessed using five performance metrics and validated against 18 newly generated SC-MRF designs. Feature importance analysis using Shapley Additive Explanations (SHAP) identified key parameters influencing EDPs, while multi-gene genetic programming (MGGP) produced explicit predictive equations to support performance-based design (PBD) of SC-MRFs. Results demonstrated that the AL-driven approach reduced the required number of training samples by 70% while maintaining at least 88% accuracy, outperforming models trained on 90 randomly selected samples. The best hybrid models achieved accuracy of 94%, 98.6%, and 99% in predicting story-level residual drift, acceleration, and maximum drift across all SC-MRF configurations. Furthermore, using MGGP-based equations, accurate performance was achieved with a minimum of 91% precision in predicting story-level EDP.