Physics-guided CNN surrogate model of viscoelastic dampers and its application in structural seismic control

Accurate characterization of the time-domain nonlinear mechanical behavior of viscoelastic dampers (VEDs) remains challenging. Traditional frequency-domain models cannot simultaneously consider loading history, strain amplitude, and ambient temperature, whereas high-fidelity time-domain models generally require complex parameter calibration and incur high computational cost. To overcome such limitations, a physics-guided convolutional neural network (CNN) surrogate model is proposed by integrating a one-dimensional CNN (1D CNN) within an equivalent linear viscoelasticity (ELV) model. A dual-branch CNN is subsequently developed to identify the ELV parameters, while the ELV model provides physical guidance through the loss function, resulting in an interpretable physics-data-driven framework. By leveraging available numerical and experimental data, the proposed model is validated from two representative energy dissipation devices (EDDs), demonstrating its ability to characterize nonlinear mechanical behavior and energy dissipation. Furthermore, the proposed model is applied to nonlinear dynamic analyses of viscoelastically damped frames under various earthquake excitations and ambient temperatures. The results show that the proposed model can accurately characterize the time-domain nonlinear mechanical behavior of two type of common EDDs, while its theoretical framework shows potential for extension to other damping devices. Structural dynamic analyses further indicate that neglecting viscoelastic nonlinearity may lead to inaccurate estimation of the energy-dissipation capacity of VEDs and introduce errors in seismic response evaluation. Moreover, ambient temperature may have a non-negligible influence on structural responses, predictions based on the nominal design temperature may deviate from the actual behavior under substantial temperature variations, which should be considered in engineering design.