Probabilistic identification method for seismic failure modes of reinforced concrete beam-column joints using Gaussian process with deep kernel

Identifying the seismic failure modes of beam-column joints (BCJs) is crucial for the safety and integrity of reinforced concrete (RC) buildings or structures withstanding seismic forces. However, traditional identification methods fail to provide any indication about the uncertainties within their predictions, which is beneficial for evaluating, interpreting and improving these predictions. This study develops a probabilistic identification method for seismic failure modes of BCJs using Gaussian process (GP) with a deep kernel. Firstly, a deep kernel architecture which reasonably describes the evolution characteristics of seismic failure modes of BCJs was proposed by transforming the deep neural network architecture into the characteristics of kernel functions. Then a probabilistic identification method for seismic failure modes of BCJs was developed by integrating the deep kernel architecture into a GP (DGP). Meanwhile, the hyper-parameters of the DGP were optimized by stochastic variational inference (SVI) strategy. Finally, the developed DGP was evaluated by comparing it with traditional shear-resistance design methods and machine learning techniques based on 289 sets of experimental data. Analysis results demonstrated the potential of the proposed method for improving the classification accuracy of traditional GPs, as well as its superiority over the prediction accuracy of traditional shear-resistance design methods and machine learning techniques. Furthermore, the proposed method also provides an efficient approach to estimate the uncertainties within their predictions for seismic failure modes of BCJs.