A machine-learning-based model for seismic performance assessment of interior slab-column connections

Abstract

Slab-column connections are the most critical components in flat plate systems under seismic actions due to their brittle failure mode. Finite element analysis (FEA) can be adopted as an extension of experimental testing to potentially lead to the development of modern design codes by providing significant information on the punching shear failure process of connections. However, the available concrete constitutive models in FEA packages fail in simulating the hysteretic response of the connections under cyclic loads. Moreover, it is shown that the available design code recommendations and empirical models are not accurate tools for assessing the seismic performance of these connections (e.g., unbalanced moment and/or drift capacity). In an effort to overcome these limitations, a dataset of previously tested experiments is compiled to predict the ultimate load and drift of the connections using the multi-variate non-linear regression model. Additionally, there is a need to develop a new constitutive material model for seismic performance assessment of the connections to capture the complete hysteretic response of the connection using numerical FEA. Consequently, a deep neural network is trained over the accumulated dataset and integrated into the OpenSees source code to determine the seismic performance of the connections. For data augmentation, a generative neural network model of autoencoders known as Generative Adversarial Networks (GANs) is constructed to generate synthetic data. The proposed material model is found to have significantly improved precision in capturing the hysteretic response of flat plates. The developed models showed higher accuracy than the design code estimations and empirical models.