ConvColPush: A convolution neural network based approach for predicting performance limit states and simplified backbone curve of RC bridge piers

Performance based seismic design (PBSD) requires identification of damage states in different structural components. Bridge piers are an integral part of the bridge system that dictate the overall performance of a bridge during a seismic event. Identifying different performance limit states require nonlinear time history analysis or performing static pushover analysis. Although the pushover analysis is simple compared to the nonlinear time history analysis, it requires adequate knowledge of nonlinear modeling and can be time consuming when a rapid assessment is required. This study implements a novel convolution neural network (CNN) based approach for performance prediction and backbone curve development of circular reinforced concrete bridge piers. At first, prototype bridge piers are modelled with OpenSees and validated with experimental results. Monte Carlo simulation is adopted to create an extensive dataset of 1000 piers with unique combination of input features. The input parameters used in this study are pier diameter ( D ), aspect ratio ( L/D ), compressive strength of concrete ( f c ’), yield strength of longitudinal steel ( f yl ), yield strength of transverse steel ( f yt ), longitudinal reinforcement ratio ( ρ l ), transverse reinforcement ratio ( ρ t ) and axial load ratio ( N y ). Different ML models such as Linear Regression (RF), Random Forest (RF), extreme gradient boosting (XGBoost), adaptive gradient boosting (AdaBoost), convolutional neural network (CNN), artificial neural network (ANN) and long short-term memory network (LSTM) are trained and tested with hyperparameter optimization being performed using grid search algorithm and 5-fold cross validation technique. The developed CNN model (ConvColPush) turns out to be the best ML model with the highest coefficient of determination ( R 2 ) value of 0.86 on test set amongst all algorithms investigated. Further validation of the developed ConvColPush is evaluated by comparing the predicted backbone curves with those obtained from experimental studies and OpenSees models. Lastly, ConvColPush is published through a graphical user interface (GUI) for practical usage by design engineers.