A comparative study of machine learning methods for assessing the compressive strength of Pozzolanic concrete

Abstract

Abstract This study aims to enhance the prediction of compressive strength in pozzolanic concrete by leveraging machine learning techniques. Unlike conventional methods relying on costly lab tests or empirical correlations, machine learning offers a more precise and efficient approach. Three models—artificial neural networks (ANNs), random forest (RF), and gradient boosting regressor (GBR)—are harnessed to develop predictive models. The dataset, comprising 482 samples, is divided randomly into 70% (337 samples) for training and 30% (145 samples) for testing. Seven input parameters related to pozzolanic material type, proportion, and mix design are utilized for model training. Model performance assessment employs metrics such as coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). Notably, the RF model outperforms others, achieving the highest R2 of 0.976 in training and 0.964 in testing, with the lowest RMSE (2.84) and MAE (2.05) in training and 7.81 and 5.89, respectively in testing, demonstrating superior predictive precision. The model’s accuracy is evaluated using the Taylor diagram. Additionally, sensitivity analysis reveals cement as the most impactful input parameter, influencing 28% of variability. The RF model’s robustness is confirmed through K-fold cross-validation, yielding an average R2 of 0.959. This study underscores the reliability and effectiveness of the RF model for forecasting pozzolanic concrete’s compressive strength, carrying implications for optimizing concrete mix and construction practices. Overall, the proposed RF model excels in efficiency and accuracy, establishing its supremacy over other algorithms in predicting pozzolanic concrete properties.