Scholarlink

<p>Boriding is an effective thermochemical surface-hardening treatment, but optimizing the resulting layer thickness typically relies on iterative experimentation. This work combines experimental analysis with explainable machine learning to provide a rapid prediction tool for two technologically important borides, Fe₂B and TiB₂. Ti-6Al-4&#xa0;V and 316L substrates were borided in a Na₂B₄O₇–SiC molten-salt bath at 850&#xa0;°C, 950&#xa0;°C, and 1000&#xa0;°C for 2–8&#xa0;h, yielding 36 thickness measurements. A feed-forward artificial neural network (ANN) trained with Bayesian-regularized Levenberg–Marquardt achieved excellent predictive performance (overall R<sup>2</sup> = 0.99, MAE &lt; 1.2&#xa0;µm), outperforming Random Forest and linear regression models. Explainable AI analyses revealed that Fe₂B growth is primarily governed by temperature (≈52% importance), while TiB₂ thickness is more sensitive to treatment time (≈58%). Local interpretable model-agnostic explanation maps confirmed these trends and highlighted a narrow high-growth regime for TiB₂ at 1000&#xa0;°C / ≥ 6&#xa0;h. The hybrid experimental–machine learning workflow provides (I) near-instantaneous thickness estimates within ± 5% of experiment, (II) identification of dominant kinetic drivers, and (III) interpretable guidance for industrial parameter optimization. Expanding the dataset and including additional alloy chemistries will further improve generalizability, paving the way for closed-loop digital control of boriding processes.</p> Graphical abstract <p></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

Predicting borided layer thickness of Fe2B and TiB2 using artificial neural network

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Mohammed Amine Khater,
  • Chaaben Arroussi,
  • Bassam Gamal Nasser Muthanna,
  • Ismail Chekalil,
  • Akshansh Mishra

摘要

Boriding is an effective thermochemical surface-hardening treatment, but optimizing the resulting layer thickness typically relies on iterative experimentation. This work combines experimental analysis with explainable machine learning to provide a rapid prediction tool for two technologically important borides, Fe₂B and TiB₂. Ti-6Al-4 V and 316L substrates were borided in a Na₂B₄O₇–SiC molten-salt bath at 850 °C, 950 °C, and 1000 °C for 2–8 h, yielding 36 thickness measurements. A feed-forward artificial neural network (ANN) trained with Bayesian-regularized Levenberg–Marquardt achieved excellent predictive performance (overall R2 = 0.99, MAE < 1.2 µm), outperforming Random Forest and linear regression models. Explainable AI analyses revealed that Fe₂B growth is primarily governed by temperature (≈52% importance), while TiB₂ thickness is more sensitive to treatment time (≈58%). Local interpretable model-agnostic explanation maps confirmed these trends and highlighted a narrow high-growth regime for TiB₂ at 1000 °C / ≥ 6 h. The hybrid experimental–machine learning workflow provides (I) near-instantaneous thickness estimates within ± 5% of experiment, (II) identification of dominant kinetic drivers, and (III) interpretable guidance for industrial parameter optimization. Expanding the dataset and including additional alloy chemistries will further improve generalizability, paving the way for closed-loop digital control of boriding processes.

Graphical abstract