Physics-informed deep neural networks for simulating S-shaped steel dampers

Abstract

Listen

Translate article

Numerical simulation that combines finite element methods and experimental data has been recognized as effective in modeling hysteretic behaviors and capturing the principle mechanical trend of passive energy dissipation devices. However, the seismic design and mechanical characteristics assessment for passive energy dissipation devices require a laborious effort and massive computational resources to tune mechanical-oriented parameters. Particularly, the potential risk of departure from the actual system is rising for the desirable seismic design of dampers due to the numerical model simplification and assumption. To eliminate the potential weakness of numerical models, this paper explores a surrogate model by implementing physics-informed deep neural networks (DNNs) to approximate hysteretic behaviors of S-shaped steel dampers. The proposed physics-informed DNNs mainly consists of recurrent neural networks (RNNs) and long short-term networks (LSTMs), which can encode the Bouc-Wen model into the direct graph and incorporate the effect of design-oriented geometry parameters. To validate the generality of the network, the optimization model was calibrated numerically and experimentally, respectively, which exhibits good performance in predicting nonlinear behaviors of different dampers with reasonable accuracy. The proposed physics-informed DNNs can be an alternative to relieve the laboriousness of the seismic design and mechanical characteristics assessment of passive energy dissipation devices.