Scholarlink

<p>This study develops a unified machine learning (ML) framework to predict the normalized hardness of friction stir welded (FSW) aluminum alloys using a consolidated dataset of several weld conditions collected from seventeen peer-reviewed sources. The input space—rotational speed, welding speed, and tool tilt angle—was first analyzed using the RReliefF algorithm, which identified rotational speed as the dominant factor governing hardness evolution across alloys. Five supervised ML models were constructed and evaluated: Gradient Boosting, Random Forest, Artificial Neural Network, Support Vector Machine, and k-Nearest Neighbors. The models were tuned through systematic hyperparameter optimization and validated using 10-fold cross-validation. Among all methods, Gradient Boosting achieved the highest predictive accuracy (R² ≈ 0.97) with stable error distributions and minimal bias across the hardness range. Model interpretability through permutation importance and SHAP-style partial effects confirmed the nonlinear interaction between heat generation, thermal exposure, and tool geometry. The results demonstrate that nonlinear ensemble methods can effectively capture thermomechanical–microstructural relationships in FSW. This work provides a transferable data-driven framework for welding parameter optimization and lays the foundation for predictive design of mechanical properties in solid-state joining.</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 framework for hardness prediction across multi-alloy friction stir welds

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Amir Reza Ansari Dezfoli,
  • Yi-Jen Huang,
  • Sanaz Hadidchi

摘要

This study develops a unified machine learning (ML) framework to predict the normalized hardness of friction stir welded (FSW) aluminum alloys using a consolidated dataset of several weld conditions collected from seventeen peer-reviewed sources. The input space—rotational speed, welding speed, and tool tilt angle—was first analyzed using the RReliefF algorithm, which identified rotational speed as the dominant factor governing hardness evolution across alloys. Five supervised ML models were constructed and evaluated: Gradient Boosting, Random Forest, Artificial Neural Network, Support Vector Machine, and k-Nearest Neighbors. The models were tuned through systematic hyperparameter optimization and validated using 10-fold cross-validation. Among all methods, Gradient Boosting achieved the highest predictive accuracy (R² ≈ 0.97) with stable error distributions and minimal bias across the hardness range. Model interpretability through permutation importance and SHAP-style partial effects confirmed the nonlinear interaction between heat generation, thermal exposure, and tool geometry. The results demonstrate that nonlinear ensemble methods can effectively capture thermomechanical–microstructural relationships in FSW. This work provides a transferable data-driven framework for welding parameter optimization and lays the foundation for predictive design of mechanical properties in solid-state joining.