4th International Conference on Earthquake Engineering

The inelastic response of beam-column connections significantly affects the seismic performances of steel frames. Accurate models of beam-column connections are important both in understanding the seismic behavior and in design. An effective approach towards developing accurate models for connections is self-learning simulation, which is presented in this paper. The latter simulation method has the unique advantage in that it can extract material models from global structural response. Selflearning simulation is based on auto-progressive algorithms that employ the principles of equilibrium and compatibility, and the self-organizing capability of artificial neural network material models. Steel beam-column connections are represented by point plastic hinges that are modeled by a new neural network representation for modeling of cyclic behavior of connections. The neural network based connection model has significant advantages over conventional models in that it can handle complex behavior due to local buckling and tearing of components. Moreover, its implementation is more efficient than conventional connection models since it does not need interaction equation and plastic potential. In this paper, the performance of the neural network based connection model is verified with experimental data and the proposed self-learning simulation methodology is demonstrated with a simulated structural testing of a two-story one bay frame with semi-rigid connections.