Structural parameter identification using physics-informed neural networks

Abstract

A parameter identification framework has been developed based on physics-informed neural networks (PINNs). Physical constraints are taken into account during the training process of a PINN, creating a grey-box running mechanism. Two information acquisition principles are proposed for training data sets and physical constraints. Specifically, finite element computation is incorporated with the uniform design to generate the minimum number of training data for PINNs. Then multivariate nonlinear regression is applied to the training data to establish the physical constraints, which are used as a rule model added to the loss function for training evaluation. This step guides the training process towards a physically or mechanically consistent solution, instead of a pure data association. Thereby the training of PINNs involves the physical governing laws, leading to a physics-informed data-driven approach. Finally, the proposed PINNs were used to identify the stiffness parameters of a laboratory-scale frame model and an actual frame structure.