Scholarlink

<p>Salt-bath and powder-packed boronizing processes were performed on AISI 304, 420, and 440C stainless steels at a temperature range of 850–1000&#xa0;°C and a soaking time of about 1–9&#xa0;h. All boride layer thicknesses were measured and analyzed using traditional diffusion kinetics and machine learning techniques, i.e., neural networks and adaptive neuro-fuzzy inference system. The six input features were boronizing time, temperature, carbon-, chromium-, nickel contents, and boronizing type. The neural network used two training algorithms, i.e., Levenberg–Marquardt and Bayesian regularization. Diffusion kinetics were found to predict the boride layer thickness with reasonable accuracy (error less than 8%). The empirical equations of boride layer thickness were presented as a function of the boronizing temperature and time. The neural networks exhibited excellent prediction tools, especially the deep neural network trained using the Bayesian regularization algorithm. An error of less than 3% was observed. Connecting weights in the deep neural network model were analyzed, and then the relative importance of input features was calculated. The adaptive neuro-fuzzy inference system exhibited good predictive performance with an error of less than 6.6%. Prediction diagrams of boride layer thickness regarding the boronizing temperature and time were finally established using diffusion kinetics, artificial-, deep neural network, and adaptive neuro-fuzzy inference system.</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 and diffusion kinetics of predicting boride layer thickness on differently boronized stainless steels

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Patiphan Juijerm,
  • Laksamee Angkurarach,
  • Patcharin Naemchanthara,
  • Oranicha Theerakiat

摘要

Salt-bath and powder-packed boronizing processes were performed on AISI 304, 420, and 440C stainless steels at a temperature range of 850–1000 °C and a soaking time of about 1–9 h. All boride layer thicknesses were measured and analyzed using traditional diffusion kinetics and machine learning techniques, i.e., neural networks and adaptive neuro-fuzzy inference system. The six input features were boronizing time, temperature, carbon-, chromium-, nickel contents, and boronizing type. The neural network used two training algorithms, i.e., Levenberg–Marquardt and Bayesian regularization. Diffusion kinetics were found to predict the boride layer thickness with reasonable accuracy (error less than 8%). The empirical equations of boride layer thickness were presented as a function of the boronizing temperature and time. The neural networks exhibited excellent prediction tools, especially the deep neural network trained using the Bayesian regularization algorithm. An error of less than 3% was observed. Connecting weights in the deep neural network model were analyzed, and then the relative importance of input features was calculated. The adaptive neuro-fuzzy inference system exhibited good predictive performance with an error of less than 6.6%. Prediction diagrams of boride layer thickness regarding the boronizing temperature and time were finally established using diffusion kinetics, artificial-, deep neural network, and adaptive neuro-fuzzy inference system.