Modeling the ductility characteristics of ultrahigh-performance fiber-reinforced concrete using a precise, computation-efficient, and comprehensive data-driven random forests-based approach

Abstract

The most relevant features of ultra-high-performance fiber-reinforced concrete (UHPFRC) are its ductility, toughness, and pos-cracking (residual or strain-hardening) parameters under direct tensile stress. Nonetheless, measuring these parameters requires complicated and delicate tests in which minor variations in test performance might render the findings invalid. This research is aimed at creating four different machine-learning algorithmic regressions for estimating the uniaxial tensile performance of UHPFRC. Hence, the first crack stress σcc, the maximum post-cracking strength σpc, and its associated strains, εcc, and εpc, were modeled through the Random Forest algorithm. The developed approaches showed a high goodness-of-fit for regression measured in the test subset, presenting R2 values of 0.894, 0.909, 0.930, and 0.911 in guesstimating the behavior of UHPFRC under direct tensile stress (viz., σcc, σpc, εcc, and εpc). The proposed models are a tool for assessing UHPFRC under ductility requirements and reducing the experimental campaign's time and costs by pre-selecting the available constituents with the proper performance on the algorithmic approaches.