EXPLAINABLE MACHINE LEARNING MODELS TO ASSESS THE BOND BEHAVIOR AND STRENGTH OF STEEL FIBERS EMBEDDED IN STEEL FIBER REINFORCED CONCRETE

An adequate bond strength between steel fibers and concrete in steel fiber reinforced concrete (SFRC) matrices is critical to achieve high toughness and tensile strength in such matrices for structural applications. In particular, predicting the equivalent bond strength is important to efficiently assess the tensile strength of these matrices and for finite element simulations of structural elements. However, predicting this bond strength is complex as it depends on many mix design variables and physical characteristics of the fiber. The currently available empirical equations rely only on the compressive strength of the concrete, limiting their accuracy. Therefore, it is important to establish an accurate method for rapid prediction of this bond strength in different fiber and matrix scenarios. To tackle this issue, this paper develops an explainable machine learning model to assess the equivalent bond strength of embedded steel fibers over a wide range of parameters for SFRC. A comprehensive dataset was assembled, consisting of 1154 pull-out tests of embedded steel fibers. Different ML algorithms, including ensemble models and neural networks, were evaluated to define an accurate predictive model. After several comparisons, the most accurate model (based on XGBoost) was selected, with a coefficient of determination over 0.82 for the testing set. Moreover, the Shapley Additive exPlanations technique was used to explain and evaluate the importance of different variables on the equivalent bond strength. To help in practical applications, a surrogate equation was also proposed to predict the equivalent bond strength of steel fibers. Both the ML model and surrogate equation proved more accurate and generalizable than existing empirical equations. The proposed models promise to accelerate SFRC development and promote its structural applications.