A novel GPR-based prediction model for cyclic backbone curves of reinforced concrete shear walls

Backbone curves are used to characterize nonlinear responses of structural elements by simplifying the cyclic force–deformation relationships. Accurate modeling of cyclic behavior can be achieved with a reliable backbone curve model. In this paper, a novel machine learning-based model is proposed to predict the backbone curve of reinforced concrete shear (structural) walls based on key wall design properties. Reported experimental responses of a detailed test database consisting of 384 reinforced concrete shear walls under cyclic loading were utilized to predict seven critical points to define the backbone curves, namely: shear at cracking point (V_cr); shear and displacement at yielding point (V_y and δ_y); and peak shear force and corresponding displacement (V_max and δ_max); and ultimate displacement and corresponding shear (V_u and δ_u). The predictive models were developed based on the Gaussian Process Regression method (GPR), which adopts a non-parametric Bayesian approach. The ability of the proposed GPR-based model to make accurate and robust estimations for the backbone curves was validated based on unseen data using a hundred random sampling procedure. The prediction accuracies (i.e., ratio of predicted/actual values) are close to 1.0, whereas the coefficient of determination (R²) values range between 0.90–0.97 for all backbone points. The proposed GPR-based backbone models are shown to reflect cyclic behavior more accurately than the traditional methods, therefore, they would serve the earthquake engineering community for better evaluation of the seismic performance of existing buildings.