Pull-out behaviour of steel fibres in concrete: experimental study and generalisable ANN modelling

Fibre-reinforced concretes (FRCs) offer flexural strength, fracture energy and durability, with their performance depending on the fibre–matrix interaction. This study develops generalisable models based on artificial neural networks (ANNs) to predict the pull-out parameters of steel fibres from a mortar matrix, incorporating the influence of data uncertainty. Ninety-six pull-out tests were performed on straight and hooked-end fibres to comprehend the influence of key features. Utilising experimental and literature data, models are trained and developed by optimising the Bayesian-regularised ANN architecture. The hyperparameters are tuned through Monte Carlo cross-validation (MCCV) by evaluating 11 250 models. Unlike the majority of models reported in literature, the proposed MCCV accounts for uncertainty in the choice of training data, as warranted by limited data, yielding reliable R2 values of 0.94 and 0.72 for predicting the bond strength and pull-out energy, respectively. Smaller R2 value (∼ 0.33) was noted for the model predicting slip at peak load owing to limited data on slip-hardening behaviour. This study provides original experimental data and reliable machine learning models to aid fibre selection and efficient design of experiments for FRCs.