Applicability of Convolutional Neural Networks for Calibration of Nonlinear Dynamic Models of Structures

Abstract

The objective of this study is to examine a machine learning (ML) framework for calibrating the parameters of analytical models of complex nonlinear structural systems where experimental data is significantly limited. Because of the high cost of large-scale structural tests, analytical models are widely used to enhance the understanding of structural performance under complex loading environments. In this study, an ML framework is proposed and evaluated for the calibration of an analytical model representing a shake table test performed on a composite column developed in OpenSees software. A large number of parameters for modeling the constitutive behavior of the concrete core, steel reinforcement, exterior composite tube, as well as the interactions between the concrete core and the tube, base fixity, and nonlinear shear deformations are included. A convolutional neural network (CNN) architecture was used to calibrate these parameters by using the lateral load, displacement, and axial load time histories as input variables. First, a synthetic dataset is generated for permutations of different model parameters. Next, four CNNs were trained to evaluate the presentation of input data in time-domain and time-frequency domain. Finally, the trained model was prompted with real experimental data and the values of peak lateral force, residual displacement, and hysteresis energy dissipation from the analytical model were compared with those from the experiment. The results show that the proposed framework is appropriate for calibration of complex nonlinear structural models when experimental data is limited.