Article
Characterization of Structural Dynamic Boundary Conditions Using Physics-Informed Neural Networks (en)
* Presenting author
Abstract:
Uncertainties arising from unknown boundary conditions are a primary source of prediction inaccuracies in vibration simulations. While inverse methods exist for estimating these boundary conditions, they always rely on a pre-validated forward model and often require a high number of model evaluations, making them impractical for computationally expensive models. This study employs physics-informed neural networks to first learn the forward model and subsequently characterize the boundary conditions in a data-driven manner. The physics-informed neural networks integrate the residual of the equations of linear elasticity into the loss function along with the training data, ensuring physically consistent predictions. The training dataset consists of noisy boundary displacement data obtained from a finite element reference solution. After training, the network learns to predict the displacement field within the structure, satisfying the Navier-Lamé equations and the given training data. Our findings demonstrate the capability of physics-informed neural networks to accurately predict the displacement field within three-dimensional structures using only boundary training data. Furthermore, the trained network enables the estimation of previously unknown boundary conditions, offering a promising data-driven approach to addressing inverse problems in the field of structural dynamics.