Flexural behaviour prediction of fibre reinforced ferrocement beams under repeated cyclic loading using artificial neural networks and experimental analysis

Abstract This research proposes a strong hybrid model in the experimental computational domain to determine and predict the flexural behavior of fibre reinforced ferrocement beams exposed to cycling loading. An initial experimental program containing six beam specimens produced critical performance characteristics on the Residual Strength Ratio (RSR) and Energy Absorption. To address the plight of the limited experimental database, a small-scale physics informed synthetic dataset of 500 samples was produced. A multi-output, feedforward Artifial Neural Network (ANN) was created and painstakingly tested to a high standard of validation that included 5-fold cross validation, blind experimental sample testing, and bootstrap sample uncertainty quantification (B = 500). Of the blind experimental samples, the ANN model was found to be universally predictive of RSR (R²=0.9209). However, remarkably lesser predictive ability was attained for the Absorption of Energy (R²= -1.5183); this discrepancy was shown to be a consequence of capture of post cracking energy dissipation mechanisms. Global sensitivity analysis by Permutation Importance of the ANN model showed that fibre volume fraction and number of mesh layers were the only parameters that accounted for 97% of the absorption of energy and 62% of the RSR. The study is the first of its kind to demonstrate the capacity of anns to predict residual strength in ferrocement. Further, it shows that predicting the energy absorption response of ferrocement requires a much larger sample size than what exists. Based on this work, researchers can now use the study results as a reference to help machine learning improve the design and optimization of ferrocement components, intended to be used in structures that exhibit repetitive loading and require other specialized materials.