Shear capacity prediction of slender reinforced concrete structures with steel fibers using machine learning

Abstract

Shear failure in reinforced concrete beams poses a critical safety issue since it may occur without any prior signs of damage in some cases. Many of the existing shear design equations for steel fiber reinforced concrete (SFRC) beams include significant uncertainty due to failure in reflecting the phenomenology of shear resistance accurately. Given these, adequate reliability evaluation of shear design provisions for SFRC beam is of high significance, and increased accuracy and minimisation of variability in the predictive model is essential. This contribution proposes machine learning (ML) based methods - Gaussian Process regression (GPR) and the Random Forest (RF) techniques - to predict the ultimate shear resistance of SFRC slender beams without stirrups. The models were developed using a database of 326 experimental SFRC slender beams obtained from previous studies, utilising 75% for model training and the remainder for testing. The performance of the proposed models was assessed by statistical comparison to experimental results and to that of the state-of-practice existing shear design models (fib Model Code 2010, German guideline, Bernat et al. model). The proposed ML-based models are in close alignment with the experimentally observed shear strength and the existing predictive models, but provide more accurate and unbiased predictions. Furthermore, the model uncertainty of the various resistance models was characterised and investigated. The ML-based models displayed the lowest bias and variability, with no significant trend with input parameters. The inconsistencies observed in the predictions by the existing shear design formulations at the variation of shear span to effective depth ratio is a major cause for concern; reliability analysis is required. Finally, partial resistance safety factors were proposed for the model uncertainty associated with the existing shear design equations.