<p>The compressive strength of concrete and composite materials is a crucial parameter for its design and evaluation. It highly influences the reliability and durability of structural members thus contributing to disaster prevention and mitigation. Traditional methods for assessing compressive strength are labor-intensive. Regression analysis stands out as a valuable statistical tool in this regard, allowing the development of predictive models based on historical data. These models establish relationships between compressive strength and factors such as material composition, curing conditions, and environmental variables. Machine learning techniques, particularly neural networks and ensemble methods like random forest and “Extreme Gradient Boosting", have gained traction for their effectiveness in predicting compressive strength across various concrete compositions and environmental conditions. This paper introduces a novel approach which proposes a tailored data splitting strategy to improve the representation of compressive strength values in training datasets, and reduce the dependence of optimal training on a probabilistic mechanism. In our study, we worked with a dataset of 396 samples. Our proposed stacking ensemble with Fuzzy C-means clustering-based splitting improved the predictive performance substantially compared to baseline models. Additionally, it combines the decision tree and linear regression models in a stacking ensemble framework, with a linear regressor serving as the meta-learner. This ensemble aims to mitigate common issues of over-fitting and under-fitting encountered in standalone models. The research enhances predictive accuracy, evidenced by a “Co-efficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>)" surpassing 0.98, a “Mean Absolute Error" of 0.6416, a “Root Mean Squared Error" of 1.0222 MPa , a “Mean Absolute Percentage Error" of 4.3076 and a “Symmetric Mean Absolute Percentage Error" of 4.3287, while also addressing the shortcomings of traditional cross-validation techniques. By enhancing the reliability and robustness of compressive strength predictions, this approach promises to support more informed decisions in structure planning and material science applications.</p>

Compressive Strength Prediction of Geopolymers Using Stacking Ensemble and Fuzzy Splitting

期刊论文
发表日期：

2025年4月23日

查看全文

Sourav Kumar Das,
Satyabrata Roy,
Srivaishnavi Yaddanapudi,
Dhruv Pradeep Chhajed

摘要

The compressive strength of concrete and composite materials is a crucial parameter for its design and evaluation. It highly influences the reliability and durability of structural members thus contributing to disaster prevention and mitigation. Traditional methods for assessing compressive strength are labor-intensive. Regression analysis stands out as a valuable statistical tool in this regard, allowing the development of predictive models based on historical data. These models establish relationships between compressive strength and factors such as material composition, curing conditions, and environmental variables. Machine learning techniques, particularly neural networks and ensemble methods like random forest and “Extreme Gradient Boosting", have gained traction for their effectiveness in predicting compressive strength across various concrete compositions and environmental conditions. This paper introduces a novel approach which proposes a tailored data splitting strategy to improve the representation of compressive strength values in training datasets, and reduce the dependence of optimal training on a probabilistic mechanism. In our study, we worked with a dataset of 396 samples. Our proposed stacking ensemble with Fuzzy C-means clustering-based splitting improved the predictive performance substantially compared to baseline models. Additionally, it combines the decision tree and linear regression models in a stacking ensemble framework, with a linear regressor serving as the meta-learner. This ensemble aims to mitigate common issues of over-fitting and under-fitting encountered in standalone models. The research enhances predictive accuracy, evidenced by a “Co-efficient of determination ( $R^2$ $R^{2}$ )" surpassing 0.98, a “Mean Absolute Error" of 0.6416, a “Root Mean Squared Error" of 1.0222 MPa , a “Mean Absolute Percentage Error" of 4.3076 and a “Symmetric Mean Absolute Percentage Error" of 4.3287, while also addressing the shortcomings of traditional cross-validation techniques. By enhancing the reliability and robustness of compressive strength predictions, this approach promises to support more informed decisions in structure planning and material science applications.