Use of optimized machine learning tool for predicting compressive strength of concrete

Abstract

Optimizing concrete mix design is essential for advancing sustainable construction practices. Conventional methods for evaluating the compressive strength (CS) of concrete, a critical mechanical property, are often time-intensive and resource-demanding. This study investigates the application of machine learning (ML) models to predict CS of concrete more efficiently, utilizing the Python interface on Google Colab. Multiple regression models have been assessed using performance metrics including mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and the coefficient of determination (R²). A stacked regression (SR) model has been developed by integrating 14 base models, with CatBoost (CB) employed as the meta-learner. The models have been trained and tested on a dataset comprising 1315 samples collected from the Concrete Laboratory at Veer Surendra Sai University of Technology (VSSUT), Burla, using an 80/20 train-test split. To enhance model performance, hyper-parameter tuning via Grid Search and validation through K-Fold cross-validation have been employed. The optimized SR-CB model has achieved superior predictive accuracy, recording an RMSE of 1.95 and an R² of 0.93. Furthermore, SHAP and LIME analyses have been conducted to interpret the feature importance and model behaviour. The model’s generalizability has been validated by predicting the CS of 21 new concrete mixes from literature, resulting in prediction errors ranging from 0.5% to 9.9% and a R² of 0.93. The findings demonstrate that the proposed stacked regression approach significantly improves prediction accuracy and robustness compared to individual models, thereby facilitating more efficient and sustainable concrete mix design with reduced dependence on conventional experimental methods.