Scholarlink

<p>Correct estimation of the fatigue strength of steels remains challenging since the relationship between chemical composition, heat-treatment history, and microstructural integrity is not only complex but also nonlinear. A physics-informed machine learning (PIML) approach is proposed to predict fatigue strength in steels in an accurate, interpretable, and uncertainty-sensitive way. Four physically motivated indices, the Tempering Index (TI), Hardening Index (HI), Diffusion–Carburizing Index (DCI), and Microstructural Integrity Index (MII), are created to encode the important heat-treatment mechanisms and microstructural influences into parsimonious, mechanism-correlated indices. Three types of features are considered, namely conventional feature, physics-informed features, and hybrid features, and are compared using multiple regression methods, such as gradient boosting and Gaussian process regression (GPR), on an industrial dataset of 437 steel samples. The hybrid representation, which adds metallurgical indices to a few alloying elements, has the best predictive performance on a held-out test set, with an R<sup>2</sup> of 0.986, a mean absolute error of 17.82&#xa0;MPa, and a root mean squared error of 24.20&#xa0;MPa. The explained factors shown in a Shapley Additive Explanations (SHAP) suggest that Diffusion–Carburizing Index and chromium content are the major factors that contributed to fatigue strength, which is in close agreement with known metallurgical principles of surface hardening and fatigue crack initiation. Moreover, the uncertainty quantification based on GPR proves that the hybrid framework yields the smallest and most consistent 95% interval confidence (±15–40&#xa0;MPa), which allows making predictions that are reliable under a wide range of processing conditions. The proposed indices improve interpretability and uncertainty-aware deployment of ML for fatigue-critical steel design.</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

Metallurgical Indices for Physics-Informed Machine Learning-Based Fatigue Strength Prediction in Steels

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Ala’a Al-Falahat

摘要

Correct estimation of the fatigue strength of steels remains challenging since the relationship between chemical composition, heat-treatment history, and microstructural integrity is not only complex but also nonlinear. A physics-informed machine learning (PIML) approach is proposed to predict fatigue strength in steels in an accurate, interpretable, and uncertainty-sensitive way. Four physically motivated indices, the Tempering Index (TI), Hardening Index (HI), Diffusion–Carburizing Index (DCI), and Microstructural Integrity Index (MII), are created to encode the important heat-treatment mechanisms and microstructural influences into parsimonious, mechanism-correlated indices. Three types of features are considered, namely conventional feature, physics-informed features, and hybrid features, and are compared using multiple regression methods, such as gradient boosting and Gaussian process regression (GPR), on an industrial dataset of 437 steel samples. The hybrid representation, which adds metallurgical indices to a few alloying elements, has the best predictive performance on a held-out test set, with an R2 of 0.986, a mean absolute error of 17.82 MPa, and a root mean squared error of 24.20 MPa. The explained factors shown in a Shapley Additive Explanations (SHAP) suggest that Diffusion–Carburizing Index and chromium content are the major factors that contributed to fatigue strength, which is in close agreement with known metallurgical principles of surface hardening and fatigue crack initiation. Moreover, the uncertainty quantification based on GPR proves that the hybrid framework yields the smallest and most consistent 95% interval confidence (±15–40 MPa), which allows making predictions that are reliable under a wide range of processing conditions. The proposed indices improve interpretability and uncertainty-aware deployment of ML for fatigue-critical steel design.