Interpretable Machine-Learning Models to Predict the Flexural Strength of Fiber-Reinforced SCM-Blended Concrete Composites

Abstract

A novel data-driven approach to predict and explain the flexural strength (FS) of fiber-reinforced supplementary cementitious material (SCM)-blended concrete composites through interpretable machine learning (ML) models is presented. First, two ensemble ML models, AdaBoost (AdB) and gradient boosting (GB), were developed based on various input parameters, such as the type and proportion of SCMs, fibers, and other materials, to predict the FS after 28 days of curing. The statistical analysis showed that the GB model outperformed the AdB model in both accuracy and error, as supported by the scatter plots and Taylor diagram. The predictions by the ML models were interpreted using Sapley additive explanations (SHAP) through bee-swarm plots to identify the relative importance and influence of each input parameter on FS at both the global and local levels. Interpretations were also made for a typical instance by force-plot to represent how the prediction was made. Furthermore, as an additional layer of interpretation, individual conditional expectation (ICE) with partial dependence plots (PDP) were also plotted to visualize the dependence between the two most and least influencing features on the FS. Based on interpretable data-driven models, the most influential parameters were the steel fiber volume and the aspect ratio of the fibers, while the least influential parameters were the maximum aggregate size and the limestone-to-binder ratio. Models based on nonlinear equations were also developed and compared with the output obtained through GB for the prediction of the FS of fiber-reinforced SCM-blended concrete composites. With this novel approach, a better understanding of how the input features affect the FS of fiber-reinforced SCM-blended concrete composites is gained and thus helps in optimizing concrete mix designs accordingly.