Reinforcing bar development length modeling using integrative support vector regression model with response surface method: New approach

Due to the non-uniformity of bond stress distribution, the full bar development length should be tested to validate the development length of the reinforcing bar embedded in concretes. The current study proposed the design of a hybrid artificial intelligence (AI) model using integration of support vector regression (SVR) coupled with response surface method (RSM) for prediction of reinforcement bar development length. Two nonlinear calibrating processes are conducted, RSM is used to connect the input data on the hardening dataset in first stage while SVR is used for determining the nonlinear relation between the hardening dataset and the output development bar stress. 534 pull-out test observations on short unit bar length are used for the modeling. Several physical dimensional properties are incorporated as input attributes for the hybrid predictive model. Stand-alone AI models, empirical formulations, and design codes are used for validation. Parametric analysis is performed to investigate the sensitivity of RSM-SVR model toward as input variables. Results evidenced the capability of the proposed RSM-RVM model to approximate the non-linear relations between physical information and bar development length. The RSM-SVR model proved to be a reliable intelligence model for bar materials design. Additionally, the model was highly consistent in the calibration of the existing design codes for more reliability. In quantitative terms, RSM-SVR model attained minimum root mean square error (RMSE = 25.60 MPa), compared with for example ACI 318-14 modeling design (306.56 MPa) or stand-alone SVR model (90.95 MPa), over the testing phase. The bar development length formulation using RSM-SVR model showed the nonlinear relationship with respect to the influence of the input variables such as transvers bars in development region, reinforcing yield stress, and concrete compressive strength in normal domain.