Physics-Informed Neural Networks for System Identification of Structural Systems with a Multiphysics Damping Model

Structural system identification is critical in resilience assessments and structural health monitoring, especially following natural hazards. Among the nonlinear structural behaviors, structural damping is a complex behavior that can be modeled as a multiphysics system wherein the structure interacts with an external thermal bath and undergoes thermalization. In this paper, we propose a novel physics-informed neural network approach for nonlinear structural system identification and demonstrate its application in multiphysics cases where the damping term is governed by a separated dynamics equation. The proposed approach, called PIDynNet, improves the estimation of the parameters of nonlinear structural systems by integrating auxiliary physics-based loss terms, one for the structural dynamics and one for the thermal transfer. These physics-based loss terms form the overall loss function in addition to a supervised data-based loss term. To ensure effective learning during the identification process, subsampling and early stopping strategies are developed. The proposed framework also has the generalization capability to predict nonlinear responses for unseen ground excitations. Two numerical experiments of nonlinear systems are conducted to demonstrate the comparative performance of PIDynNet.