Scholarlink

<p>Yield strength (YS) is a critical parameter for the engineering design and safety assessment of austenitic stainless steels. However, strong collinearity among chemical composition features and the lack of physical constraints in black-box models often lead to unstable feature attribution and physically inconsistent extrapolation. To address these issues, a physics-informed parallel neural network framework is proposed for predicting the YS of 316 stainless steel. The framework employs a parallel merged neural network (PMNN), in which chemical composition and grain size are modeled independently through separate channels, followed by high-level feature fusion. Furthermore, monotonicity constraints and a Hall–Petch-based physical loss term are incorporated to construct a physics-informed PMNN (PI-PMNN), thereby enhancing prediction robustness and physical consistency. The results show that PMNN achieves an R<sup>2</sup> of 0.94 with a root mean square error (RMSE) of 8.94&#xa0;MPa. With the introduction of physical constraints, PI-PMNN further improves the R<sup>2</sup> to 0.98 and reduces the RMSE to 5.04&#xa0;MPa, corresponding to a 43.6% reduction in prediction error. SHAP analysis demonstrates improved stability of feature attribution, while univariate analysis confirms grain-size responses consistent with the Hall–Petch relationship. Overall, the proposed framework effectively balances predictive accuracy, interpretability stability, and physical consistency, providing a reliable tool for YS prediction of 316 stainless steel.</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

Prediction of Yield Strength in 316 Stainless Steel Using a Physics-Informed Parallel Network

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Hongyan Duan,
  • Ke Xu,
  • Xiao Li,
  • HongXia Jiang

摘要

Yield strength (YS) is a critical parameter for the engineering design and safety assessment of austenitic stainless steels. However, strong collinearity among chemical composition features and the lack of physical constraints in black-box models often lead to unstable feature attribution and physically inconsistent extrapolation. To address these issues, a physics-informed parallel neural network framework is proposed for predicting the YS of 316 stainless steel. The framework employs a parallel merged neural network (PMNN), in which chemical composition and grain size are modeled independently through separate channels, followed by high-level feature fusion. Furthermore, monotonicity constraints and a Hall–Petch-based physical loss term are incorporated to construct a physics-informed PMNN (PI-PMNN), thereby enhancing prediction robustness and physical consistency. The results show that PMNN achieves an R2 of 0.94 with a root mean square error (RMSE) of 8.94 MPa. With the introduction of physical constraints, PI-PMNN further improves the R2 to 0.98 and reduces the RMSE to 5.04 MPa, corresponding to a 43.6% reduction in prediction error. SHAP analysis demonstrates improved stability of feature attribution, while univariate analysis confirms grain-size responses consistent with the Hall–Petch relationship. Overall, the proposed framework effectively balances predictive accuracy, interpretability stability, and physical consistency, providing a reliable tool for YS prediction of 316 stainless steel.