Scholarlink

<p>This study presents a comprehensive experimental and AI-based investigation into the three-point bending behavior of aramid and glass fiber-reinforced hybrid composites, considering different EKabor-II nanoparticle reinforcement levels (0 wt%, 0.5 wt%, and 1 wt%). Artificial Neural Networks (ANN), Classification Learner, and Regression Learner algorithms are comparatively applied for the first time to predict the mechanical responses of composite panels with millimeter-level precision. The ANN model, trained using Levenberg–Marquardt with 5 neurons and a learning rate of 0.013, achieved a validation MSE of 0.678&#xa0;MPa and <i>R</i><sup><i>2</i></sup> of 0.9997. Meanwhile, the Gaussian Process Regression (GPR) method produced outstanding results (RMSE = 0.089&#xa0;MPa, <i>R</i><sup><i>2</i></sup> = 0.9999). Hyperparameter optimization using the CMA-ES algorithm eliminated the need for manual trial-and-error, objectively identifying the optimal ANN configuration and enhancing global search capability and generalization reliability. Five-fold cross-validation and 95% confidence intervals (RMSE = 0.75 ± 0.83&#xa0;MPa; <i>R</i><sup><i>2</i></sup> = 0.9869 ± 0.0039) demonstrate consistent performance beyond randomness. SHAP-based explainability analysis revealed that compressive load (55% contribution) and test duration (20%) dominantly influence flexural stress, enabling causal interpretation of the model. Edge-case analysis under extreme configurations (Wt = 0, Wt = 1, and maximum compressive load) confirmed prediction deviations within ± 5&#xa0;MPa, ensuring safety margins. This holistic approach significantly contributes to accelerating computational materials design and establishing reliable infrastructures for Industry 4.0, digital twin, and sustainable (eco-composite) applications in materials science and engineering.</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

Optimizing multilayer perceptron (MLP) hyperparameters via covariance matrix adaptation evolution strategy (CMA-ES) for predicting composite bending behavior

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Fatma Bakal Gumus,
  • Hayri Yildirim

摘要

This study presents a comprehensive experimental and AI-based investigation into the three-point bending behavior of aramid and glass fiber-reinforced hybrid composites, considering different EKabor-II nanoparticle reinforcement levels (0 wt%, 0.5 wt%, and 1 wt%). Artificial Neural Networks (ANN), Classification Learner, and Regression Learner algorithms are comparatively applied for the first time to predict the mechanical responses of composite panels with millimeter-level precision. The ANN model, trained using Levenberg–Marquardt with 5 neurons and a learning rate of 0.013, achieved a validation MSE of 0.678 MPa and R2 of 0.9997. Meanwhile, the Gaussian Process Regression (GPR) method produced outstanding results (RMSE = 0.089 MPa, R2 = 0.9999). Hyperparameter optimization using the CMA-ES algorithm eliminated the need for manual trial-and-error, objectively identifying the optimal ANN configuration and enhancing global search capability and generalization reliability. Five-fold cross-validation and 95% confidence intervals (RMSE = 0.75 ± 0.83 MPa; R2 = 0.9869 ± 0.0039) demonstrate consistent performance beyond randomness. SHAP-based explainability analysis revealed that compressive load (55% contribution) and test duration (20%) dominantly influence flexural stress, enabling causal interpretation of the model. Edge-case analysis under extreme configurations (Wt = 0, Wt = 1, and maximum compressive load) confirmed prediction deviations within ± 5 MPa, ensuring safety margins. This holistic approach significantly contributes to accelerating computational materials design and establishing reliable infrastructures for Industry 4.0, digital twin, and sustainable (eco-composite) applications in materials science and engineering.