Machine learning unveils the complex nonlinearity of concrete materials’ uniaxial compressive strength

Abstract

Uniaxial compressive strength (CS), a mechanical property that depends on the type of fine and coarse aggregates, the water-cement (w/c) ratio, age, the volume of admixtures, etc., is difficult to estimate for concrete materials. Hence, an investigation on the application of machine learning (ML) techniques to predict the UCS of concrete precisely and affordably is described here. Different machine learning algorithms, including ensemble models like Classification and regression trees (CART), XGBoost (XGB), Bagging (Bagg), AdaBoost (AdaBo) and Random Forest Regression (RF), and non-ensemble models like Ridge regression (RR), Partial least square (PLS), K-Nearest Neighbors algorithm (KNN) were used. A total of 1206 experimental results datasets were collected. The key input parameters for the soft computing method, which produces the CS of concrete as the output, are composed of nine different elements. The quantity of cement, fine and coarse aggregate, amount of water, plasticizer, slump, and days of curing are some of the major input components. By comparing the ML-derived results with the experimental findings using various performance indices, the study demonstrates that the XGBoost model effectively approximates the CS of concrete materials with reliability and robustness. Furthermore, the research includes sensitivity analysis based on the developed optimal XGBoost model, revealing the influence of different concrete mix parameters on CS and highlighting the strongly nonlinear nature of concrete materials.