An explainable machine learning model for encompassing the mechanical strength of polymer-modified concrete

Abstract

Polymer-modified concrete (PMC) is an advanced building material with more excellent durability, tensile strength, adhesion, and lesser susceptibility to chemical degradation. Recent developments in machine learning (ML) have shown that prediction of compressive strength (CS) of PMC key input factors needed to obtain an optimized mix design are among the areas of applicability of ML. This study used eight machine learning models, which are Decision Tree, Support Vector Machine, K-Nearest Neighbors, Bagging Regression, XG-Boost, Ada-Boost, Linear Regression, Gradient Boosting to predict compressive strength and perform SHAP (Shapley additive explanation) analysis. These hybrid predictive PMC models were developed using a wide-ranging dataset of 382 experimental data points compiled from the literature. A SHAP interaction plot was also used to show how each feature affected predictions on the model outputs. As highlighted in the results, hybrid models had significantly higher performance than conventional models, and the XG-Boost and decision tree model had the highest accuracy. In particular, the XG-Boost and decision tree model reached R² scores of 0.987 for training and 0.577 for testing, proving its remarkable prediction ability for PMC compressive strength. The SHAP analysis confirmed that coarse aggregate, cement, and SCMs had the most significant influence on CS, with all other variables contributing lower values. The Partial Dependence Plots (PDP) analysis allowed a relatively simple interpretation of the contribution of individual inputs to the CS predictions. These results are useful for construction purposes and provide engineers and builders with first-hand knowledge and insight into the importance of individual components on PMC development and performance.