Scholarlink

<p>This study integrates Scheffe’s simplex lattice (SSL) mixture design methodology with machine learning (ML) algorithms to optimize sustainable concrete mix formulations. A mathematical model was developed using 41 independent mixture ratios, from which 246 concrete specimens (150&#xa0;mm cubes) were cast, cured, and evaluated for 28-day compressive strength. The dataset was divided into training (73% of the total dataset) and testing (27% of the total dataset) subsets. Experimental compressive strength data derived from the SSL design served as the baseline for developing and validating four regression-based predictive models: relevance vector regression (RVR), elastic net regression (ENR), Gaussian process regression (GPR), and multilinear regression (MLR). Model performance varied significantly across the algorithms. ENR achieved the best predictive accuracy, with a coefficient of determination (R<sup>2</sup>) of 0.9674 (training) and 0.9235 (testing), coupled with low errors (mean absolute error, MAE: 0.6212/0.9388, root mean squared error, RMSE: 0.8193/1.1732). RVR exhibited moderate predictive power and reasonable generalization, recording an R<sup>2</sup> of 0.6405, MAE of 1.3705, and RMSE of 1.8947. MLR produced modest results with an R<sup>2</sup> of 0.7256, MAE of 0.8706, and RMSE of 1.2069. Conversely, GPR underperformed severely, with an R<sup>2</sup> of 0.0086 (training) and –0.2834 (testing), along with high error levels (MAE: 1.6956/2.3599, RMSE: 2.249/2.6245), indicating poor convergence with experimental data. The findings underscore the importance of algorithm selection and hyperparameter optimization in ML-driven concrete mix design. The integration of SSL-based experimental mixture design with regression-based predictive modeling establishes a robust and computationally efficient framework for optimizing sustainable cementitious systems. The superior performance of ENR and RVR demonstrates their suitability for capturing the complex, nonlinear, and multivariate interactions governing compressive strength, thereby advancing eco-efficient, data-driven approaches to sustainable construction 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

Modelling the compressive strength of economical and sustainable concrete with machine learning techniques

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Terngu Thaddeus Tarzomon,
  • Chijioke Christopher Ikeagwuani,
  • Onuegbu Okoronkwo Ugwu,
  • Michael Chizobam Uwaezuoke

摘要

This study integrates Scheffe’s simplex lattice (SSL) mixture design methodology with machine learning (ML) algorithms to optimize sustainable concrete mix formulations. A mathematical model was developed using 41 independent mixture ratios, from which 246 concrete specimens (150 mm cubes) were cast, cured, and evaluated for 28-day compressive strength. The dataset was divided into training (73% of the total dataset) and testing (27% of the total dataset) subsets. Experimental compressive strength data derived from the SSL design served as the baseline for developing and validating four regression-based predictive models: relevance vector regression (RVR), elastic net regression (ENR), Gaussian process regression (GPR), and multilinear regression (MLR). Model performance varied significantly across the algorithms. ENR achieved the best predictive accuracy, with a coefficient of determination (R2) of 0.9674 (training) and 0.9235 (testing), coupled with low errors (mean absolute error, MAE: 0.6212/0.9388, root mean squared error, RMSE: 0.8193/1.1732). RVR exhibited moderate predictive power and reasonable generalization, recording an R2 of 0.6405, MAE of 1.3705, and RMSE of 1.8947. MLR produced modest results with an R2 of 0.7256, MAE of 0.8706, and RMSE of 1.2069. Conversely, GPR underperformed severely, with an R2 of 0.0086 (training) and –0.2834 (testing), along with high error levels (MAE: 1.6956/2.3599, RMSE: 2.249/2.6245), indicating poor convergence with experimental data. The findings underscore the importance of algorithm selection and hyperparameter optimization in ML-driven concrete mix design. The integration of SSL-based experimental mixture design with regression-based predictive modeling establishes a robust and computationally efficient framework for optimizing sustainable cementitious systems. The superior performance of ENR and RVR demonstrates their suitability for capturing the complex, nonlinear, and multivariate interactions governing compressive strength, thereby advancing eco-efficient, data-driven approaches to sustainable construction material design.