Scholarlink

<p>The accurate prediction of the mechanical behavior in aluminum-silicon carbide (Al-SiC) composites is essential for optimizing processing conditions and ensuring a&#xa0;reliable performance in engineering applications. In recent years, machine learning has emerged as a&#xa0;powerful tool to capture complex, nonlinear interactions among processing variables. In this work, two regression frameworks—artificial neural network (ANN) and extreme gradient boosting (XGBoost)—were systematically developed and optimized using GridSearchCV. The ANN baseline showed poor predictive capability with an 𝑅<sup>2</sup> value of 0.214, a&#xa0;mean absolute error (MAE) of 22.62, and a&#xa0;root mean square error (RMSE) of 27.48. In contrast, XGBoost achieved a&#xa0;much higher accuracy with 𝑅<sup>2</sup> = 0.716, MAE = 13.30, and RMSE = 16.52, confirming its suitability as the base learner. In order to address the need for a&#xa0;lightweight and interpretable framework, a&#xa0;surrogate model was adopted wherein an ANN was trained to replicate the predictions of the XGBoost model. The surrogate ANN demonstrated strong fidelity, achieving 𝑅<sup>2</sup> = 0.796, MAE = 9.20, and RMSE = 12.62 when compared against XGBoost outputs. These results indicate that ANN surrogates can effectively approximate complex gradient boosting models, offering a&#xa0;computationally efficient and scalable pathway for real-time prediction and decision support in composite material 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

Surrogate-assisted machine learning framework for accurate prediction of mechanical behavior in Al/SiC composites

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Jagadish Lolugu,
  • Mahesh Kaza,
  • Venkateswara Reddy Kunduru,
  • K. Venkateswara Reddy

摘要

The accurate prediction of the mechanical behavior in aluminum-silicon carbide (Al-SiC) composites is essential for optimizing processing conditions and ensuring a reliable performance in engineering applications. In recent years, machine learning has emerged as a powerful tool to capture complex, nonlinear interactions among processing variables. In this work, two regression frameworks—artificial neural network (ANN) and extreme gradient boosting (XGBoost)—were systematically developed and optimized using GridSearchCV. The ANN baseline showed poor predictive capability with an 𝑅2 value of 0.214, a mean absolute error (MAE) of 22.62, and a root mean square error (RMSE) of 27.48. In contrast, XGBoost achieved a much higher accuracy with 𝑅2 = 0.716, MAE = 13.30, and RMSE = 16.52, confirming its suitability as the base learner. In order to address the need for a lightweight and interpretable framework, a surrogate model was adopted wherein an ANN was trained to replicate the predictions of the XGBoost model. The surrogate ANN demonstrated strong fidelity, achieving 𝑅2 = 0.796, MAE = 9.20, and RMSE = 12.62 when compared against XGBoost outputs. These results indicate that ANN surrogates can effectively approximate complex gradient boosting models, offering a computationally efficient and scalable pathway for real-time prediction and decision support in composite material design.