TRANSFER LEARNING SCHEME TO EFFICIENTLY ASSESS THE FLEXURAL STRENGTH OF FIBER-REINFORCED CONCRETE WITH STEEL AND POLYPROPYLENE FIBERS

Characterizing the flexural strength of fiber-reinforced concrete (FRC) is fundamental to classifying and evaluating the material's adequacy for structural applications. This property is significant in the design of structures because it directly relates to the ability to resist the loads applied and to control cracking under service loads. However, there are some challenges in rapidly predicting this flexural strength as it depends on mix design and fiber characteristic parameters. Existing empirical prediction equations include limited variables that make them produce unrealistic results since they do not capture the full complexity of the behavior. Consequently, a more efficient approach is necessary to predict the flexural strength of FRC. To tackle this issue, this paper develops explainable models to assess the flexural strength of FRC with steel and polypropylene fibers, using machine learning (ML) and transfer learning (TL) techniques. Two datasets were collected, one containing 4-point bending tests for Steel Fiber Reinforced Concrete (SFRC) and another for Polypropylene Fiber Reinforced Concrete (PFRC). Different ML algorithms were evaluated to define an accurate and efficient predictive model for SFRC. However, owing to the limited data from experimental tests of PFRC, a transfer learning approach was used to transfer the learned knowledge from the SFRC dataset to make predictions for PFRC, reducing the need for extensive retraining. The proposed models proved to be highly accurate and flexible for rapid predictions in the studied materials. The explainability of the final models is discussed through Partial Dependence Plots (PDP), which provide insights into the contribution of variables to the flexural strength, which can accelerate the characterization process of FRC, obtaining design recommendations for its use in structural elements.