A novel combined experimental-machine learning approach to estimate the probabilistic capacity of RC beams with spatially correlated rebar corrosion in transverse and longitudinal directions

Chloride-induced corrosion of tensile rebars in reinforced concrete (RC) structures causes cracking in the concrete surface along corroded rebars. The width of these cracks could provide valuable information for estimating the amount of steel weight loss inside concrete beams. However, an experimental investigation revealed that the distribution of cracks in RC beams with multiple rebars was affected not only by pressure from the corrosion expansion of the corresponding rebar but also from that of adjacent rebars. This leads to a highly complex nonlinear relationship between crack width and amount of steel corrosion. In this study, a novel combined experimental-machine learning approach is developed to estimate steel corrosion distributions in RC beams. This procedure applies generative adversarial networks (GANs) to consider the effects of spatially correlated rebar corrosion in transverse and longitudinal directions. A pix2pix network is trained by the distributions of a dataset of steel weight loss that is generated based on random field theory with the statistical parameters identified using the experimental evidence and the distributions of a dataset of corrosion crack widths constructed using finite element (FE) analysis. Subsequently, the probability density function (PDF) of the flexural capacity for corroded RC beams is obtained using Monte Carlo-based FE analysis. A case study investigating the effect of the distributions of observed crack widths on the PDF of the flexural capacity for aging RC beams with spatially correlated rebar corrosion in transverse and longitudinal directions is presented.