DNN-based estimation of flexural strength of horizontally curved bracing system

Horizontally curved (HC) I-girder bridges are highly complex due to their initial curvature, which induces significant nonlinear behavior at the strength limit state. To accurately estimate flexural strength of HC bridges, nonlinear analysis is typically required, and it involves multiple simulations to identify the most economical bridge alignment. In this study, we proposed an alternative method using deep neural networks (DNNs) to streamline the bridge design process based on nonlinear structural response data. To achieve this, data were collected in terms of the flexural strength (f_bu) and lateral stress (f_l), incorporating geometrically nonlinear FEA with imperfections (GNIA) to capture the nonlinear effects of the HC bridges. To construct the DNN architecture, 17 design parameters were selected as input variables, which are closely linked to the LTB strength of beam-bracing system for straight bridges and flexural-warping torsional behavior of HC-bridge systems. Additionally, two outputs (f_bu, f_l) were chosen as target variables. The best prediction models were evaluated using 20 % of the test datasets, along with additional FEA datasets excluded from the training. Finally, this study proposes the most effective DNN prediction model outlining a straightforward application process utilizing the developed DNN architecture. Furthermore, it was found that the selected inputs, which are closely related to the LTB limit state, show a good correlation with the selected target variables (f_bu, f_l). In addition, it was concluded that the proposed DNN models can predict flexural strength and lateral flange stress of the HC bridges under various loading condition showing with reasonable accuracy.