Self-learning simulation method for inverse nonlinear modeling of cyclic behavior of connections

This paper presents an improved self-learning simulation and its application to modeling of 'cyclic' behavior of the connections from the results of structural testing. Unlike other inverse modeling approaches such as parameter optimization methods, the proposed method requires no prior knowledge about the behavior and model. It can extract the cyclic connection models by imposing experimental measurements to the dual finite element models as boundary conditions. A new algorithmic tangent formulation during the self-learning simulation has been proposed to improve performances of the self-learning simulation. Moreover, a new neural network (NN) based hysteretic material model is utilized to expedite learning of the cyclic behavior and it is integrated into the improved self-learning simulation method. To guide a practical implementation of the self-learning simulation, numerical procedures are also presented in detail. Using both synthetic and actual experimental data, the self-learning simulation method has proven to be a reliable method to extract nonlinear cyclic models of the local connections from the global response of the framed structures.