Scholarlink

<p>Welding-induced residual stresses significantly affect the structural performance of steel members; however, their evaluation remains computationally demanding. This study presents a physics-informed, data-efficient machine learning framework for the rapid prediction of residual stress distributions in welded steel I-girders. The framework decouples thermal process prediction from residual stresses estimation. In Stage I, Gaussian Process Regression is used to predict key thermal descriptors, including interpass cooling temperature, peak temperature, cooling rate, and heat-affected zone width, based on geometric, material, and process parameters. These descriptors are then used in Stage II, where a gradient-boosted decision tree model predicts tensile and compressive stress peaks as well as transition lengths. Mechanical admissibility is ensured through a physics-constrained optimization procedure that enforces material yield limits and cross-sectional equilibrium. In addition, a Δ-learning calibration module is introduced to compensate for systematic discrepancies between simulation results and experimental measurements. The framework is trained on 128 thermomechanical finite element simulations and validated against independent numerical cases and a full-scale experimental girder. The results demonstrate high prediction accuracy, substantial reduction in computational time, and the ability to reproduce experimentally observed structural behavior, supporting the application of the proposed framework in design and fabrication optimization.</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 hybrid decoupled machine learning framework with physical constraints for predicting welding-induced residual stresses in steel girders

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月17日

查看全文
  • Jingwen Zhang,
  • Wenbo Huang,
  • Zhipeng Ni,
  • Yao Zhang

摘要

Welding-induced residual stresses significantly affect the structural performance of steel members; however, their evaluation remains computationally demanding. This study presents a physics-informed, data-efficient machine learning framework for the rapid prediction of residual stress distributions in welded steel I-girders. The framework decouples thermal process prediction from residual stresses estimation. In Stage I, Gaussian Process Regression is used to predict key thermal descriptors, including interpass cooling temperature, peak temperature, cooling rate, and heat-affected zone width, based on geometric, material, and process parameters. These descriptors are then used in Stage II, where a gradient-boosted decision tree model predicts tensile and compressive stress peaks as well as transition lengths. Mechanical admissibility is ensured through a physics-constrained optimization procedure that enforces material yield limits and cross-sectional equilibrium. In addition, a Δ-learning calibration module is introduced to compensate for systematic discrepancies between simulation results and experimental measurements. The framework is trained on 128 thermomechanical finite element simulations and validated against independent numerical cases and a full-scale experimental girder. The results demonstrate high prediction accuracy, substantial reduction in computational time, and the ability to reproduce experimentally observed structural behavior, supporting the application of the proposed framework in design and fabrication optimization.