Development of a Standalone Application for Accurate and User-Friendly Prediction of Concrete Compressive Strength Using Ensemble Machine Learning

Authors: Tufail Mabood
Journal:  Emirati Journal of Civil Engineering and Applications
Volume: Vol 4 Issue 1
Keywords: Concrete compressive strength, stacked ensemble, meta,
learning, machine learning, feature importance, cross validation

Abstract:

Concrete compressive strength (CCS) is a key factor that affects the structural service life of structure. Laboratory testing is time consuming, costly, and restricted in applicability for the mix designs. Machine learning (ML) has emerged as a potential alternative to laboratory testing, based upon understanding the non-linear interaction of the variables for the mix design and CCS. This study builds a stacked ensemble (meta-learning) approach for eight distinct machine learning algorithms: Linear Regression, Tree, Random Forest, Gradient Boosting, AdaBoost, K Nearest Neighbors, Support Vector Machine, and Neural Networks. The UCI benchmark dataset (1,030 examples) with eight features (cement, blast furnace slag, fly ash, water, superplasticizer, coarse/fine aggregates, age) and CCS as the target was analyzed. 70/30 splits for train and test sets and multi, level k-fold cross, validation (2–20 folds) were employed for robustness. Analysis of model performance was mainly carried out using R² permutation, based feature importances and one way ANOVA for the categorical variable age. The stacked model resulted in the best overall R²=0.890 (20, fold CV) compared to the best singles’ performances (random forest: R²=0.878, gradient boosting: R²=0.874). Graphs for the predicted and actual CCS confirmed a very close fit (variations < 7 MPa). The cross validated model’s overall features’ importances were dominated by cement, age, and water, which was confirmed using the resulting ANOVA for age influence. As a spin-off, a convenient standalone software for the proposed framework also exists for real-time CCS strength predictions. The standalone application and trained models developed in this study will be made publicly available upon paper acceptance at: https://github.com/tufailmab/ccs-ensemble-predictor