Predicting load capacity of shear walls using SVR–RSM model

Abstract Accurate prediction of the shear capacity of reinforced concrete shear walls (RCSW) is essential for the wind and seismic design of buildings. However, due to the diverse structural configurations, multitude of load scenarios, and highly nonlinear relations between the design parameters and the shear load capacity, this prediction is very complex. Existing pertinent design code provisions such as the American Concrete Institute ACI-318 and the Eurocode rely on empirical expressions that have various limitations and attain low predictive accuracy. Hence, in this paper, we pioneer a novel hybrid intelligent model to predict the ultimate shear capacity of RCSW. The support vector regression (SVR) and response surface model (RSM) were coupled based on two calibrating strategies in a novel hybrid modeling approach called RSM–SVR. The accuracy, tendency and uncertainty of the proposed SVR–RSM model along with that of three existing empirical relations and two design code provisions were assessed using various statistical metrics based on a comprehensive experimental database retrieved from the open literature. The existing design codes and empirical models were found to be inflicted with high variability and did not capture the influence of the key design parameters on the shear capacity in a robust and rational manner. Conversely, it is shown that the proposed RSM–SVR modeling approach achieved superior accurate predictions for the shear strength of RCSW. The proposed RSM–SVRmodel enhanced RMSE for the training (testing) dataset by 510% (150%) compared to the Baghi et al. model, 550% (190%) compared to the ACI 318-14 design code, 530% (155%) compared to the Chandra et al. model, 320% (145%) compared to the RSM model, and 450% (90%) compared to the SVR model. The novel approach also better captured the influence of the key design parameters, demonstrating robust tendency and much lower uncertainty. Thus, the proposed novel model could be harvested in intelligent generative design and for the enhancement of pertinent provisions in design codes. The proposed method achieves outstanding performance, while maintaining superior computational efficiency and low run time.