Machine learning-based prediction for residual bearing capacity and failure modes of rectangular corroded RC columns

Abstract

Chloride-induced corrosion of steel reinforcement is likely to change the failure modes of reinforced concrete (RC) columns in marine environments under the combined action of vertical and seismic load. This study was aimed to predict the residual horizontal bearing capacity and failure modes of corroded RC columns based on machine learning (ML) methods. First, a database (163 flexural failure, 26 flexural-shear failure, and 35 shear failure) of rectangular corroded RC columns was established according to predefined selection criteria. Then, 6 efficient ML algorithms, including support vector machine, artificial neural network, k-nearest neighbors, convolutional neural network, random forest, and categorical boosting, were employed to establish predictive models for evaluating the residual bearing capacity and failure modes of corroded RC columns. Based on the SHapley Additive exPlanations (SHAP) method, the importance of input characteristic parameters was ranked and how the input parameters make decisions to obtain the final output results explained. Finally, influences of stirrup corrosion level (ηt), shear-span ratio (λ), and axial load ratio (n) were investigated regarding the residual bearing capacity and failure mode evolution of corroded RC columns. The results show that the proposed ML models can effectively predict the residual bearing capacity and failure modes of corroded RC columns and exhibit great potential for describing the relationship between input variables and output results.