Scholarlink

<p>The objective of this investigation is to integrate a <i>machine learning</i> (ML) approach with experimental results to predict the 28&#xa0;day compressive strength of high-performance concrete, which has been designed for use as a structural layer in pavements. This high-performance concrete has been developed by modifying it with <i>silica fume</i> (SF) and <i>graphene oxide</i> (GO). In order to create a predictive data set from experimental testing of all mixes, a comprehensive experimental program that included 750 individual test specimens was conducted. The experimental program tested the control (CC), SF-only (CS1-CS4), GO-only (CG1-CG4), and the optimized factorial mixes of SF-GO (CS1G1-CS4G4) were used to develop the predictive data set. Testing results showed a marked improvement in the mechanical behavior when the optimal dosages of SF = 7% and GO = 0.15%, which resulted in a hybrid mix of CS3G3 with a compressive strength increase of 29.30% compared to the control concrete (77&#xa0;MPa vs. 59.7&#xa0;MPa). Also, there were improvements in flexural strength, split-tensile strength, elastic modulus, reduced water absorption, reduced permeable voids, improved fatigue life, and durability (acid and sulfate). Using a train-test ratio of 80:20, ten fold cross-validation to assess all models; eight supervised ML algorithms were tested (Linear Regression, Decision Trees, Support Vector Regressor, AdaBoost Regressor, Random Forest, XGBoost, and Gradient Boosting) with voting algorithm. Performance of each model was determined by four engineering indexes (A10, A20, PI, IA, OBJ) in addition to common statistical evaluation (RMSE, MAE, MSE, R<sup>2</sup> and CV%). Ensemble models outperformed baseline learners, with <i>Random Forest</i> achieving the highest predictive accuracy (RMSE = 2.47&#xa0;MPa, R<sup>2</sup> score = 0.776, A10 ≈ 0.995, A20 = 1.0), closely followed by <i>XGBoos</i>t and <i>Gradient Boosting</i>. Feature-importance ranking consistently identified SF, GO and OPC as the primary predictors governing strength development, reflecting micro and nano-scale synergistic enhancement mechanisms. Violin-plot dispersion, predicted-versus-actual alignment and estimator-sensitivity analysis confirmed the robustness, stability and reproducibility of ensemble-based modelling. Overall, the findings demonstrate that ML- based predictive frameworks, when supported by the experimentally consistent datasets and engineering validation metrics, can provide reliable strength forecasts suitable for performance-based <i>optimization</i> of high-performance nano-concrete mixes.</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-driven modeling of strength in high-performance graphene oxide induced nano-concrete composites

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • B. Ramesh Rao,
  • Shriram Marathe,
  • Abhay Sridhar,
  • Akarsh Pattan Kotrabasappa

摘要

The objective of this investigation is to integrate a machine learning (ML) approach with experimental results to predict the 28 day compressive strength of high-performance concrete, which has been designed for use as a structural layer in pavements. This high-performance concrete has been developed by modifying it with silica fume (SF) and graphene oxide (GO). In order to create a predictive data set from experimental testing of all mixes, a comprehensive experimental program that included 750 individual test specimens was conducted. The experimental program tested the control (CC), SF-only (CS1-CS4), GO-only (CG1-CG4), and the optimized factorial mixes of SF-GO (CS1G1-CS4G4) were used to develop the predictive data set. Testing results showed a marked improvement in the mechanical behavior when the optimal dosages of SF = 7% and GO = 0.15%, which resulted in a hybrid mix of CS3G3 with a compressive strength increase of 29.30% compared to the control concrete (77 MPa vs. 59.7 MPa). Also, there were improvements in flexural strength, split-tensile strength, elastic modulus, reduced water absorption, reduced permeable voids, improved fatigue life, and durability (acid and sulfate). Using a train-test ratio of 80:20, ten fold cross-validation to assess all models; eight supervised ML algorithms were tested (Linear Regression, Decision Trees, Support Vector Regressor, AdaBoost Regressor, Random Forest, XGBoost, and Gradient Boosting) with voting algorithm. Performance of each model was determined by four engineering indexes (A10, A20, PI, IA, OBJ) in addition to common statistical evaluation (RMSE, MAE, MSE, R2 and CV%). Ensemble models outperformed baseline learners, with Random Forest achieving the highest predictive accuracy (RMSE = 2.47 MPa, R2 score = 0.776, A10 ≈ 0.995, A20 = 1.0), closely followed by XGBoost and Gradient Boosting. Feature-importance ranking consistently identified SF, GO and OPC as the primary predictors governing strength development, reflecting micro and nano-scale synergistic enhancement mechanisms. Violin-plot dispersion, predicted-versus-actual alignment and estimator-sensitivity analysis confirmed the robustness, stability and reproducibility of ensemble-based modelling. Overall, the findings demonstrate that ML- based predictive frameworks, when supported by the experimentally consistent datasets and engineering validation metrics, can provide reliable strength forecasts suitable for performance-based optimization of high-performance nano-concrete mixes.