Physics-guided nonlinear response prediction of RC Columns using partial simulation data and machine learning

This paper presents a physics-guided method for accurate nonlinear response prediction of Reinforced Concrete (RC) columns by combining partial physical data with Machine Learning (ML) predictions. The overall method employs the Extended Kalman Filter (EKF) to fuse partial response data from a baseline simulation model with ML predictions, effectively correcting and enhancing the predicted nonlinear response of RC columns. ML algorithms are initially employed to predict the force-deformation response of the specimens using a database of structural and geometric properties for 202 RC columns under quasi-static cyclic loading. In parallel, partial physical data is derived from numerical simulations obtained from two scenarios, ranging from low-resolution to highly detailed modeling. The EKF is then used to merge the ML predictions with the partial physical data, enabling iterative correction and refinement of the predicted nonlinear response. The proposed dual-stage model enhances the accuracy of the ML algorithms prediction for the entire regions of the backbone curve, from 0.81 to 0.88, 0.91, and 0.93, respectively, for low-resolution, median-resolution, and detailed modeling. To validate the proposed method, a case study was conducted by Finite Element (FE) modeling of a cyclically loaded RC column at varying resolutions. The accuracy of the force-deformation response has been enhanced through the integration of partial FE data via the EKF implementation, resulting in an average improvement of 10% across the entire backbone curve.