Current AISC-LRFD code requires that the moment-rotation characteristics ofconnections be known. Moreover, it requires that these characteristics be incorporated inthe analysis and member design under factored loads (AISC, 2001). Conventionalmodeling approaches to improve the prediction of cyclic behavior starts with a choice of a phenomenological model followed by calibration of the model parameters. However, not only is the improvement limited due to inherent limitations of this approach, but also test results indicate a large variability in load-carrying capacity under earthquake loading.In this research, a new neural network (NN) based cyclic material model is applied toinelastic hysteretic behavior of connections. In the proposed model, two energy-based internal variables are introduced to expedite the learning of hysteretic behavior of materials or structural components. The model has significant advantages over conventional models in that it can handle complex behavior due to local buckling andtearing of connecting elements. Moreover, its numerical implementation is more efficient than the conventional models since it does not need an interaction equation and a plastic potential. A new approach based on a self-learning simulation algorithm is used to characterize the hysteretic behavior of the connections from structural tests. The proposed approach is verified by applying it to both synthetic and experimental examples. For its practical application in semi-rigid connections, design variables are included as inputs to the model through a physical principle based module. The extended model also gives reasonable predictions under earthquake loads even when it is presented with new geometrical properties and loading scenario as well.
展开▼
