Abstract:To achieve rapid lateral capacity prediction of corroded RC columns without specific failure modes under seismic action, residual lateral capacity of corroded rectangular RC columns has been investigated based on machine learning (ML) methods. First, the input parameters for the ML model were determined, the dataset of corroded rectangular RC column was reorganized and correlations among the input parameters were analyzed. Then, six typical ML algorithms, including SVM, KNN, ANN, CNN, RF and CatBoost, were selected to establish the ML models for lateral capacity of corroded RC columns, and the hyper-parameters of ML models were optimized and determined using grid search or trial-and-error methods as well as 5-fold cross-validation. Finally, the predictive accuracy of the six ML models was evaluated based on four widely-used regression evaluation metrics (EMA, ERMS, EMAP and R2) and testing set, and the importance ranking of input parameters and parameter sensitivity analysis were carried out based on the excellent ML models. Results show that the predictive performance of neural network models is generally better than that of ensemble models or other single models, in which ANN model is more satisfactory, and the mean and standard deviation of the ratio of its predicted result to experimental value are 1.01 and 0.14, respectively. Shear-span ratio has the highest and significant effect on the lateral capacity of corroded rectangular RC columns, with an importance of up to 35%. The lateral capacity of rectangular RC columns decreases with the increase of shear-span ratio, the ratio of stirrup spacing to section effective height and corrosion level, but increases at first and then tends to stabilize or decrease with the increase of axial load ratio.